diff --git a/docs/superpowers/plans/2026-06-23-model-ranking-plan.md b/docs/superpowers/plans/2026-06-23-model-ranking-plan.md
new file mode 100644
index 0000000..8fb166f
--- /dev/null
+++ b/docs/superpowers/plans/2026-06-23-model-ranking-plan.md
@@ -0,0 +1,194 @@
+# Model Ranking Feature Implementation Plan
+
+**Date:** 2026-06-23
+**Feature:** Model KVCache Size Ranking
+**Branch:** feat/model-ranking
+
+## Overview
+
+Implement a model ranking feature that allows users to compare different models by their KV cache size requirements. The feature will be available both in the web UI (calculator.html) and as a CLI command.
+
+## Context
+
+From brainstorming session:
+- **Platform:** Both web UI and CLI
+- **Metrics:** KV bytes per token AND total KV cache for configurable sequence lengths
+- **Model scope:** Existing presets only (Qwen3-27B, Llama3.1-70B, Qwen2.5-72B, Mistral-7B, Llama3.1-8B, DeepSeek-V3-671B)
+- **UI presentation:** New section in calculator
+- **Deployment config:** Include tp_size/pp_size to show per-rank KV cache size
+
+## Implementation Tasks
+
+### Task 1: Extract Model Registry
+
+**Goal:** Create a single source of truth for model specifications
+
+**Steps:**
+1. Extract model profiles from PRESETS into a separate MODELS object
+2. Each model entry includes: name, family, model_profile, description
+3. Update PRESETS to reference MODELS instead of duplicating model_profile
+4. Verify calculator still works with existing presets
+
+**Files:** `website/calculator.html`
+
+**Verification:** Open calculator.html, select different presets, verify model fields populate correctly
+
+---
+
+### Task 2: Add Ranking Calculation Functions (JavaScript)
+
+**Goal:** Implement ranking logic in the web calculator
+
+**Steps:**
+1. Add `calculateModelRanking()` function that:
+   - Takes sequence_length as input
+   - Iterates through all models in MODELS
+   - Calculates for each model:
+     - `kv_bytes_per_token` (using kvBytesPerToken)
+     - `kv_bytes_per_token_per_rank` (using kvBytesPerTokenPerRank)
+     - `total_kv_bytes` = kv_bytes_per_token * sequence_length
+     - `total_kv_bytes_per_rank` = kv_bytes_per_token_per_rank * sequence_length
+     - `total_kv_gb` = total_kv_bytes / BYTES_PER_GB
+   - Returns sorted array (default: by total_kv_gb descending)
+2. Add `formatBytes()` helper for human-readable byte formatting
+3. Add `renderRankingTable(rankingData, sortBy)` function that:
+   - Renders HTML table with columns: Rank, Model, KV/Token, KV/Token/Rank, Total KV (X tokens), Total GB
+   - Supports sorting by clicking column headers
+   - Highlights currently selected model
+
+**Files:** `website/calculator.html`
+
+**Verification:** Open browser console, call `calculateModelRanking(8192)`, verify output structure
+
+---
+
+### Task 3: Add Model Ranking UI Section
+
+**Goal:** Create user-facing ranking interface in calculator
+
+**Steps:**
+1. Add new section after "Model Preset" card:
+   ```html
+   <div class="card">
+     <div class="card-title">Model KVCache Ranking</div>
+     <div class="card-body">
+       <label>Sequence Length (tokens)</label>
+       <input type="number" id="ranking-sequence-length" value="8192" />
+       <button id="calculate-ranking-btn">Calculate Ranking</button>
+       <div id="ranking-table-container"></div>
+     </div>
+   </div>
+   ```
+2. Style the section with CSS (add to existing <style> block)
+3. Wire up event listeners:
+   - "Calculate Ranking" button triggers calculateModelRanking() with input value
+   - Render results in #ranking-table-container
+4. Add auto-calculate on page load with default 8192 tokens
+5. Add sort functionality to table headers
+
+**Files:** `website/calculator.html`
+
+**Verification:**
+- Open calculator.html
+- Verify ranking section appears
+- Click "Calculate Ranking" button
+- Verify table renders with all models
+- Click column headers to sort
+- Change sequence length and recalculate
+
+---
+
+### Task 4: Implement CLI Command
+
+**Goal:** Add `rank-models` subcommand to Python CLI
+
+**Steps:**
+1. Create `src/kvcache_upper_bound/cli/rank_models.py`:
+   - Add `add_rank_models_parser(subparsers)` function
+   - Arguments:
+     - `--sequence-length` (required, int, can specify multiple)
+     - `--format` (csv or markdown, default: markdown)
+     - `--output` (optional file path, default: stdout)
+   - Main function `run_rank_models(args)`:
+     - Load model profiles from configs/ or hardcoded registry
+     - For each model, calculate metrics for each sequence length
+     - Format output as markdown table or CSV
+     - Write to output or stdout
+2. Register command in `src/kvcache_upper_bound/__main__.py`
+3. Add model registry to `src/kvcache_upper_bound/core/models.py`:
+   - `MODEL_REGISTRY` dict with preset model profiles
+   - Helper function `get_all_model_profiles()`
+
+**Files:**
+- `src/kvcache_upper_bound/cli/rank_models.py` (new)
+- `src/kvcache_upper_bound/__main__.py`
+- `src/kvcache_upper_bound/core/models.py`
+
+**Verification:**
+```bash
+python -m kvcache_upper_bound rank-models --sequence-length 4096 --sequence-length 8192
+python -m kvcache_upper_bound rank-models --sequence-length 32768 --format csv
+```
+
+---
+
+### Task 5: Test and Polish
+
+**Goal:** Ensure feature works end-to-end and matches existing code quality
+
+**Steps:**
+1. Web UI testing:
+   - Test with different sequence lengths (1024, 8192, 32768, 131072)
+   - Verify sorting works for all columns
+   - Test with different model presets selected (should highlight in table)
+   - Check responsive design on mobile viewport
+2. CLI testing:
+   - Test markdown output format
+   - Test CSV output format
+   - Test with multiple sequence lengths
+   - Test output to file
+3. Edge cases:
+   - Very small sequence length (1 token)
+   - Very large sequence length (1M tokens)
+   - Models with kv_cache_layer_count set vs null
+4. Documentation:
+   - Add section to README.md about rank-models command
+   - Update README.zh.md with Chinese translation
+   - Add inline comments explaining ranking formula
+
+**Files:**
+- `README.md`
+- `README.zh.md`
+- `website/calculator.html` (polish CSS/UX)
+
+**Verification:**
+- Run through all test scenarios above
+- Verify no console errors
+- Verify CLI help text is clear
+
+---
+
+## Success Criteria
+
+1. ✅ Web UI shows ranking table with all preset models
+2. ✅ Users can input custom sequence lengths and see total KV cache
+3. ✅ Table is sortable by any column
+4. ✅ CLI command outputs ranking in both markdown and CSV formats
+5. ✅ Metrics include both per-token and total KV cache sizes
+6. ✅ Deployment config (tp_size/pp_size) is reflected in per-rank calculations
+7. ✅ Currently selected model is highlighted in the ranking table
+8. ✅ Documentation updated in both README files
+
+## Technical Notes
+
+- KV bytes per token formula: `2 * layers * n_kv_heads * head_dim * dtype_bytes`
+- Use `kv_cache_layer_count` if set, otherwise `n_layers`
+- Per-rank calculation divides by `tp_size * pp_size`
+- Total KV = per-token * sequence_length
+- Sort default: by total KV cache descending (largest first)
+
+## Dependencies
+
+- No new Python dependencies
+- No new JavaScript dependencies (uses vanilla JS)
+- Reuses existing model profile structure and calculation functions
diff --git a/src/kvcache_upper_bound/cli/main.py b/src/kvcache_upper_bound/cli/main.py
index 73128ba..223a4e5 100644
--- a/src/kvcache_upper_bound/cli/main.py
+++ b/src/kvcache_upper_bound/cli/main.py
@@ -185,6 +185,29 @@ def main() -> int:
         help="Allow replay-only synthetic hash_ids when benchmark records do not provide them",
     )
 
+    rank_models_parser = subparsers.add_parser(
+        "rank-models",
+        help="Rank models by KV cache size requirements for given sequence lengths",
+    )
+    rank_models_parser.add_argument(
+        "--sequence-length",
+        type=int,
+        action="append",
+        required=True,
+        help="Sequence length in tokens (can specify multiple times)",
+    )
+    rank_models_parser.add_argument(
+        "--format",
+        choices=("markdown", "csv"),
+        default="markdown",
+        help="Output format (default: markdown)",
+    )
+    rank_models_parser.add_argument(
+        "--output",
+        default=None,
+        help="Output file path (default: stdout)",
+    )
+
     args = parser.parse_args()
     if args.command == "list-datasets":
         return _run_list_datasets()
@@ -200,6 +223,8 @@ def main() -> int:
         return _run_convert_conversation_dataset(args)
     if args.command == "convert-benchmark-results":
         return _run_convert_benchmark_results(args)
+    if args.command == "rank-models":
+        return _run_rank_models(args)
     raise ValueError(f"unsupported command: {args.command}")
 
 
@@ -487,6 +512,203 @@ def _run_convert_benchmark_results(args: argparse.Namespace) -> int:
     return 0
 
 
+MODEL_REGISTRY = {
+    "qwen3-27b": {
+        "name": "Qwen3-27B",
+        "family": "Qwen",
+        "n_layers": 64,
+        "kv_cache_layer_count": 16,
+        "n_kv_heads": 4,
+        "head_dim": 256,
+        "dtype_bytes": 2,
+        "tp_size": 8,
+        "pp_size": 1,
+    },
+    "llama3.1-70b": {
+        "name": "Llama 3.1 70B",
+        "family": "Llama",
+        "n_layers": 80,
+        "kv_cache_layer_count": None,
+        "n_kv_heads": 8,
+        "head_dim": 128,
+        "dtype_bytes": 2,
+        "tp_size": 8,
+        "pp_size": 1,
+    },
+    "qwen2.5-72b": {
+        "name": "Qwen2.5-72B",
+        "family": "Qwen",
+        "n_layers": 80,
+        "kv_cache_layer_count": None,
+        "n_kv_heads": 8,
+        "head_dim": 128,
+        "dtype_bytes": 2,
+        "tp_size": 8,
+        "pp_size": 1,
+    },
+    "mistral-7b": {
+        "name": "Mistral 7B",
+        "family": "Mistral",
+        "n_layers": 32,
+        "kv_cache_layer_count": None,
+        "n_kv_heads": 8,
+        "head_dim": 128,
+        "dtype_bytes": 2,
+        "tp_size": 1,
+        "pp_size": 1,
+    },
+    "llama3.1-8b": {
+        "name": "Llama 3.1 8B",
+        "family": "Llama",
+        "n_layers": 32,
+        "kv_cache_layer_count": None,
+        "n_kv_heads": 8,
+        "head_dim": 128,
+        "dtype_bytes": 2,
+        "tp_size": 1,
+        "pp_size": 1,
+    },
+    "deepseek-v3-671b": {
+        "name": "DeepSeek-V3-671B",
+        "family": "DeepSeek",
+        "n_layers": 61,
+        "kv_cache_layer_count": None,
+        "n_kv_heads": 1,
+        "head_dim": 512,
+        "dtype_bytes": 2,
+        "tp_size": 8,
+        "pp_size": 1,
+    },
+}
+
+
+def _run_rank_models(args: argparse.Namespace) -> int:
+    sequence_lengths = args.sequence_length
+    output_format = args.format
+
+    rankings = []
+    for model_id, model in MODEL_REGISTRY.items():
+        layers = model["kv_cache_layer_count"] if model["kv_cache_layer_count"] is not None else model["n_layers"]
+        kv_per_token = 2 * layers * model["n_kv_heads"] * model["head_dim"] * model["dtype_bytes"]
+        shard_factor = model["tp_size"] * model["pp_size"]
+        kv_per_token_per_rank = kv_per_token // shard_factor
+
+        entry = {
+            "model_id": model_id,
+            "name": model["name"],
+            "family": model["family"],
+            "kv_bytes_per_token": kv_per_token,
+            "kv_bytes_per_token_per_rank": kv_per_token_per_rank,
+            "layers": layers,
+            "n_kv_heads": model["n_kv_heads"],
+            "head_dim": model["head_dim"],
+            "tp_size": model["tp_size"],
+            "pp_size": model["pp_size"],
+        }
+
+        for seq_len in sequence_lengths:
+            total_bytes = kv_per_token * seq_len
+            total_bytes_per_rank = kv_per_token_per_rank * seq_len
+            entry[f"total_kv_gb_{seq_len}"] = total_bytes / (1024**3)
+            entry[f"total_kv_gb_{seq_len}_per_rank"] = total_bytes_per_rank / (1024**3)
+
+        rankings.append(entry)
+
+    rankings.sort(key=lambda x: x.get(f"total_kv_gb_{sequence_lengths[0]}", 0), reverse=True)
+
+    if output_format == "csv":
+        output = _format_rankings_csv(rankings, sequence_lengths)
+    else:
+        output = _format_rankings_markdown(rankings, sequence_lengths)
+
+    if args.output:
+        Path(args.output).write_text(output, encoding="utf-8")
+    else:
+        print(output)
+
+    return 0
+
+
+def _format_bytes(size_bytes: int) -> str:
+    """Format bytes as human-readable string."""
+    if size_bytes < 1024:
+        return f"{size_bytes} B"
+    elif size_bytes < 1024**2:
+        return f"{size_bytes / 1024:.1f} KB"
+    elif size_bytes < 1024**3:
+        return f"{size_bytes / (1024**2):.1f} MB"
+    else:
+        return f"{size_bytes / (1024**3):.2f} GB"
+
+
+def _format_rankings_markdown(rankings: list[dict], sequence_lengths: list[int]) -> str:
+    """Format rankings as markdown table."""
+    lines = []
+    lines.append("# Model KV Cache Ranking")
+    lines.append("")
+
+    for seq_len in sequence_lengths:
+        lines.append(f"## Sequence Length: {seq_len:,} tokens")
+        lines.append("")
+
+        header = "| Model | Family | KV/Token | KV/Token/Rank | Total KV | Total KV/Rank | Layers | KV Heads | Head Dim | TP | PP |"
+        separator = "|-------|--------|----------|---------------|----------|---------------|--------|----------|----------|----|----|"
+        lines.append(header)
+        lines.append(separator)
+
+        for rank, entry in enumerate(rankings, 1):
+            total_gb = entry.get(f"total_kv_gb_{seq_len}", 0)
+            total_gb_rank = entry.get(f"total_kv_gb_{seq_len}_per_rank", 0)
+            kv_token = _format_bytes(entry["kv_bytes_per_token"])
+            kv_token_rank = _format_bytes(entry["kv_bytes_per_token_per_rank"])
+
+            lines.append(
+                f"| {entry['name']} | {entry['family']} | {kv_token} | {kv_token_rank} | "
+                f"{total_gb:.2f} GB | {total_gb_rank:.2f} GB | {entry['layers']} | "
+                f"{entry['n_kv_heads']} | {entry['head_dim']} | {entry['tp_size']} | {entry['pp_size']} |"
+            )
+
+        lines.append("")
+
+    return "\n".join(lines)
+
+
+def _format_rankings_csv(rankings: list[dict], sequence_lengths: list[int]) -> str:
+    """Format rankings as CSV."""
+    lines = []
+
+    base_headers = ["model_id", "name", "family", "kv_bytes_per_token", "kv_bytes_per_token_per_rank",
+                    "layers", "n_kv_heads", "head_dim", "tp_size", "pp_size"]
+    seq_headers = []
+    for seq_len in sequence_lengths:
+        seq_headers.append(f"total_kv_gb_{seq_len}")
+        seq_headers.append(f"total_kv_gb_{seq_len}_per_rank")
+
+    lines.append(",".join(base_headers + seq_headers))
+
+    for entry in rankings:
+        base_values = [
+            entry["model_id"],
+            entry["name"],
+            entry["family"],
+            str(entry["kv_bytes_per_token"]),
+            str(entry["kv_bytes_per_token_per_rank"]),
+            str(entry["layers"]),
+            str(entry["n_kv_heads"]),
+            str(entry["head_dim"]),
+            str(entry["tp_size"]),
+            str(entry["pp_size"]),
+        ]
+        seq_values = []
+        for seq_len in sequence_lengths:
+            seq_values.append(f"{entry.get(f'total_kv_gb_{seq_len}', 0):.6f}")
+            seq_values.append(f"{entry.get(f'total_kv_gb_{seq_len}_per_rank', 0):.6f}")
+
+        lines.append(",".join(base_values + seq_values))
+
+    return "\n".join(lines)
+
+
 def _build_analysis_metadata_payload(
     trace: str,
     config_path: str,
diff --git a/src/kvcache_upper_bound/heuristic/multi_agent.py b/src/kvcache_upper_bound/heuristic/multi_agent.py
index 1028b8c..53e824f 100644
--- a/src/kvcache_upper_bound/heuristic/multi_agent.py
+++ b/src/kvcache_upper_bound/heuristic/multi_agent.py
@@ -88,6 +88,34 @@ def content_hit_rate(self) -> float:
         hit_tokens = self.shared_prefix_tokens + self.average_reusable_private_tokens_per_agent()
         return min(1.0, hit_tokens / average_request_tokens)
 
+    def per_turn_reusable_private_tokens(self) -> list[float]:
+        """Return reusable private tokens at each turn step."""
+        return [
+            min(self.private_window_tokens, step * self.avg_new_tokens_per_turn)
+            for step in range(self.avg_turns_per_session)
+        ]
+
+    def per_turn_content_hit_rates(self) -> list[float]:
+        """Return content hit rate at each turn step."""
+        rates = []
+        for step in range(self.avg_turns_per_session):
+            private_reuse = min(self.private_window_tokens, step * self.avg_new_tokens_per_turn)
+            request_tokens = self.shared_prefix_tokens + self.avg_new_tokens_per_turn + private_reuse
+            if request_tokens <= 0:
+                rates.append(0.0)
+            else:
+                hit = self.shared_prefix_tokens + private_reuse
+                rates.append(min(1.0, hit / request_tokens))
+        return rates
+
+    def per_turn_working_set_tokens(self) -> list[float]:
+        """Return cumulative working set tokens at each turn step."""
+        results = []
+        for step in range(self.avg_turns_per_session):
+            private_reuse = min(self.private_window_tokens, step * self.avg_new_tokens_per_turn)
+            results.append(self.shared_prefix_tokens + private_reuse + self.avg_new_tokens_per_turn)
+        return results
+
     def strict_saturation_capacity_tokens(self) -> float:
         return self.shared_prefix_tokens + self.total_private_working_set_tokens()
 
diff --git a/tests/test_multi_turn_accuracy.py b/tests/test_multi_turn_accuracy.py
new file mode 100644
index 0000000..ee10dfd
--- /dev/null
+++ b/tests/test_multi_turn_accuracy.py
@@ -0,0 +1,49 @@
+from kvcache_upper_bound.heuristic.multi_agent import (
+    MultiAgentHeuristicConfig,
+    CurveShapeConfig,
+    PolicyEfficiency,
+)
+
+
+def test_per_turn_reusable_private_tokens():
+    config = MultiAgentHeuristicConfig(
+        concurrent_agents=100,
+        shared_prefix_tokens=4096,
+        avg_new_tokens_per_turn=2048,
+        avg_turns_per_session=8,
+        private_window_tokens=32768,
+    )
+    per_turn = config.per_turn_reusable_private_tokens()
+    assert len(per_turn) == 8
+    assert per_turn[0] == 0
+    assert per_turn[1] == 2048
+    assert per_turn[7] == min(32768, 7 * 2048)
+
+
+def test_per_turn_content_hit_rates():
+    config = MultiAgentHeuristicConfig(
+        concurrent_agents=100,
+        shared_prefix_tokens=4096,
+        avg_new_tokens_per_turn=2048,
+        avg_turns_per_session=8,
+        private_window_tokens=32768,
+    )
+    rates = config.per_turn_content_hit_rates()
+    assert len(rates) == 8
+    assert rates[0] < rates[7]
+    # Turn 0: hit=4096, req=4096+2048+0=6144, rate=4096/6144=0.667
+    assert abs(rates[0] - 4096 / 6144) < 0.001
+
+
+def test_per_turn_working_set():
+    config = MultiAgentHeuristicConfig(
+        concurrent_agents=100,
+        shared_prefix_tokens=4096,
+        avg_new_tokens_per_turn=2048,
+        avg_turns_per_session=4,
+        private_window_tokens=32768,
+    )
+    wst = config.per_turn_working_set_tokens()
+    assert len(wst) == 4
+    assert wst[0] == 4096 + 0 + 2048  # shared + 0 private + new
+    assert wst[1] == 4096 + 2048 + 2048
diff --git a/tests/test_vectors_heuristic.json b/tests/test_vectors_heuristic.json
index bb3562e..03386d5 100644
--- a/tests/test_vectors_heuristic.json
+++ b/tests/test_vectors_heuristic.json
@@ -3,9 +3,9 @@
     "name": "qwen3-27b-8xh20",
     "summary": {
       "content_hit_rate": 0.8181818181818182,
-      "average_request_tokens": 22528.0,
-      "total_working_set_tokens": 29364224.0,
-      "avg_reusable_private_tokens_per_agent": 14336.0,
+      "average_request_tokens": 22528,
+      "total_working_set_tokens": 29364224,
+      "avg_reusable_private_tokens_per_agent": 14336,
       "hbm_kv_gb_per_card": 89.53163421526551,
       "strict_prefix_saturation_capacity_gb": 1792.25,
       "lru_like_saturation_capacity_gb": 2986.916666666667,
@@ -70,15 +70,35 @@
         "lru_min_cards": 8,
         "lru_min_machines": 1
       }
+    ],
+    "per_turn_private_tokens": [
+      0,
+      4096,
+      8192,
+      12288,
+      16384,
+      20480,
+      24576,
+      28672
+    ],
+    "per_turn_hit_rates": [
+      0.5,
+      0.6666666666666666,
+      0.75,
+      0.8,
+      0.8333333333333334,
+      0.8571428571428571,
+      0.875,
+      0.8888888888888888
     ]
   },
   {
     "name": "qwen3-27b-1xh20",
     "summary": {
       "content_hit_rate": 0.8181818181818182,
-      "average_request_tokens": 22528.0,
-      "total_working_set_tokens": 29364224.0,
-      "avg_reusable_private_tokens_per_agent": 14336.0,
+      "average_request_tokens": 22528,
+      "total_working_set_tokens": 29364224,
+      "avg_reusable_private_tokens_per_agent": 14336,
       "hbm_kv_gb_per_card": 44.2530737221241,
       "strict_prefix_saturation_capacity_gb": 1792.25,
       "lru_like_saturation_capacity_gb": 2986.916666666667,
@@ -143,15 +163,35 @@
         "lru_min_cards": 3,
         "lru_min_machines": 3
       }
+    ],
+    "per_turn_private_tokens": [
+      0,
+      4096,
+      8192,
+      12288,
+      16384,
+      20480,
+      24576,
+      28672
+    ],
+    "per_turn_hit_rates": [
+      0.5,
+      0.6666666666666666,
+      0.75,
+      0.8,
+      0.8333333333333334,
+      0.8571428571428571,
+      0.875,
+      0.8888888888888888
     ]
   },
   {
     "name": "edge-linear-small",
     "summary": {
       "content_hit_rate": 0.7777777777777778,
-      "average_request_tokens": 4608.0,
-      "total_working_set_tokens": 788480.0,
-      "avg_reusable_private_tokens_per_agent": 1536.0,
+      "average_request_tokens": 4608,
+      "total_working_set_tokens": 788480,
+      "avg_reusable_private_tokens_per_agent": 1536,
       "hbm_kv_gb_per_card": 66.9614839553833,
       "strict_prefix_saturation_capacity_gb": 96.25,
       "lru_like_saturation_capacity_gb": 137.39285714285714,
@@ -197,15 +237,27 @@
         "lru_min_cards": 4,
         "lru_min_machines": 4
       }
+    ],
+    "per_turn_private_tokens": [
+      0,
+      1024,
+      2048,
+      3072
+    ],
+    "per_turn_hit_rates": [
+      0.6666666666666666,
+      0.75,
+      0.8,
+      0.8333333333333334
     ]
   },
   {
     "name": "edge-saturated",
     "summary": {
       "content_hit_rate": 0.8333333333333334,
-      "average_request_tokens": 12288.0,
-      "total_working_set_tokens": 139264.0,
-      "avg_reusable_private_tokens_per_agent": 2048.0,
+      "average_request_tokens": 12288,
+      "total_working_set_tokens": 139264,
+      "avg_reusable_private_tokens_per_agent": 2048,
       "hbm_kv_gb_per_card": 89.53163421526551,
       "strict_prefix_saturation_capacity_gb": 8.5,
       "lru_like_saturation_capacity_gb": 10.5,
@@ -251,6 +303,16 @@
         "lru_min_cards": 16,
         "lru_min_machines": 2
       }
+    ],
+    "per_turn_private_tokens": [
+      0,
+      2048,
+      4096
+    ],
+    "per_turn_hit_rates": [
+      0.8,
+      0.8333333333333334,
+      0.8571428571428571
     ]
   }
 ]
diff --git a/tests/validate_js_parity.mjs b/tests/validate_js_parity.mjs
index c44fd8b..d7f6a16 100644
--- a/tests/validate_js_parity.mjs
+++ b/tests/validate_js_parity.mjs
@@ -156,6 +156,28 @@ for (const vector of vectors) {
     if (er.lru_min_cards !== null) check(`${prefix}.lru_min_cards`, ar.lruMinCards, er.lru_min_cards, INT_TOL);
     if (er.lru_min_machines !== null) check(`${prefix}.lru_min_machines`, ar.lruMinMachines, er.lru_min_machines, INT_TOL);
   }
+
+  // Per-turn analysis checks
+  if (vector.per_turn_hit_rates && vector.per_turn_private_tokens) {
+    const heuristicCfg = config.heuristic_multi_agent;
+    const createHeuristic = new Script("(function(cfg) { return new MultiAgentHeuristic(cfg); })");
+    const factory = createHeuristic.runInContext(ctx);
+    const heuristicObj = factory(heuristicCfg);
+    const jsHitRates = heuristicObj.perTurnContentHitRates();
+    const jsPrivateTokens = heuristicObj.perTurnReusablePrivateTokens();
+    const expectedHitRates = vector.per_turn_hit_rates;
+    const expectedPrivateTokens = vector.per_turn_private_tokens;
+
+    check("per_turn_hit_rates.length", jsHitRates.length, expectedHitRates.length, INT_TOL);
+    check("per_turn_private_tokens.length", jsPrivateTokens.length, expectedPrivateTokens.length, INT_TOL);
+
+    for (let t = 0; t < expectedHitRates.length; t++) {
+      check(`per_turn_hit_rates[${t}]`, jsHitRates[t], expectedHitRates[t], RATE_TOL);
+    }
+    for (let t = 0; t < expectedPrivateTokens.length; t++) {
+      check(`per_turn_private_tokens[${t}]`, jsPrivateTokens[t], expectedPrivateTokens[t], TOKEN_TOL);
+    }
+  }
 }
 
 console.log("\n" + "=".repeat(50));
diff --git a/website/calculator.html b/website/calculator.html
index 4f105d8..04711d7 100644
--- a/website/calculator.html
+++ b/website/calculator.html
@@ -404,6 +404,18 @@
         </div>
       </div>
 
+      <!-- Model KVCache Ranking -->
+      <div class="card">
+        <div class="card-title">Model KVCache Ranking</div>
+        <div class="card-body">
+          <div class="form-field">
+            <label>Sequence Length (tokens)</label>
+            <input type="number" id="ranking-sequence-length" value="8192" min="1" step="1024" />
+          </div>
+          <div id="ranking-table-container" style="overflow-x:auto;"></div>
+        </div>
+      </div>
+
       <!-- Deployment Preset -->
       <div class="card">
         <div class="card-title">Configuration Preset</div>
@@ -595,6 +607,15 @@
         <div id="sensitivity-container"></div>
       </div>
 
+      <!-- Per-Turn Analysis Chart -->
+      <div class="card" id="per-turn-chart-card" style="display:none">
+        <div class="card-title">Per-Turn Analysis</div>
+        <p style="color:var(--text-muted);font-size:.85rem">How cache hit rate evolves across conversation turns</p>
+        <div class="chart-container">
+          <canvas id="perTurnChart" height="250"></canvas>
+        </div>
+      </div>
+
       <!-- Results Table -->
       <div class="card">
         <div class="card-title">Tier Analysis</div>
@@ -630,14 +651,14 @@
     { id: "llama3.1-70b", family: "Llama", name: "Llama 3.1 70B", params: { n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 70600000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "GQA" },
     { id: "llama3.1-8b", family: "Llama", name: "Llama 3.1 8B", params: { n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 8000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "GQA" },
     { id: "llama3.3-70b", family: "Llama", name: "Llama 3.3 70B", params: { n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 70600000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "GQA" },
-    { id: "deepseek-v2-236b", family: "DeepSeek", name: "DeepSeek-V2 236B", params: { n_layers: 60, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 512, dtype_bytes: 2, parameter_count: 236000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE (compressed KV: effective n_kv_heads=16, head_dim=512)" },
-    { id: "deepseek-v3-671b", family: "DeepSeek", name: "DeepSeek-V3 671B", params: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE (compressed KV: effective n_kv_heads=16, head_dim=512)" },
+    { id: "deepseek-v2-236b", family: "DeepSeek", name: "DeepSeek-V2 236B", params: { n_layers: 60, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 236000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE (compressed KV latent: kv_lora_rank=512)" },
+    { id: "deepseek-v3-671b", family: "DeepSeek", name: "DeepSeek-V3 671B", params: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE (compressed KV latent: kv_lora_rank=512)" },
     { id: "mixtral-8x22b", family: "Mistral", name: "Mixtral 8x22B", params: { n_layers: 56, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 141000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MoE, GQA" },
     { id: "mistral-7b", family: "Mistral", name: "Mistral 7B", params: { n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 7200000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "GQA" },
     { id: "glm4-9b", family: "GLM", name: "GLM-4 9B", params: { n_layers: 40, kv_cache_layer_count: null, n_kv_heads: 2, head_dim: 128, dtype_bytes: 2, parameter_count: 9400000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MQA" },
     { id: "yi1.5-34b", family: "Yi", name: "Yi-1.5 34B", params: { n_layers: 60, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 34400000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "GQA" },
-    { id: "deepseek-v3.1-685b", family: "DeepSeek", name: "DeepSeek-V3.1 685B (V3-0324)", params: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 512, dtype_bytes: 2, parameter_count: 685000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE 256E/8A (compressed KV: kv_lora_rank=512)" },
-    { id: "kimi-k2-1t", family: "Kimi", name: "Kimi K2 1T", params: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 512, dtype_bytes: 2, parameter_count: 1000000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE 384E/8A (DeepSeek-V3 architecture)" },
+    { id: "deepseek-v3.1-685b", family: "DeepSeek", name: "DeepSeek-V3.1 685B (V3-0324)", params: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 685000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE 256E/8A (compressed KV latent: kv_lora_rank=512)" },
+    { id: "kimi-k2-1t", family: "Kimi", name: "Kimi K2 1T", params: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 1000000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "MLA + MoE 384E/8A (DeepSeek-V3 architecture, compressed KV latent: kv_lora_rank=512)" },
     { id: "gemma3-27b", family: "Gemma", name: "Gemma 3 27B", params: { n_layers: 62, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 128, dtype_bytes: 2, parameter_count: 27000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "Hybrid sliding window (1024) + global attention (every 6th layer)" },
     { id: "gemma3-12b", family: "Gemma", name: "Gemma 3 12B", params: { n_layers: 48, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 256, dtype_bytes: 2, parameter_count: 12000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "Hybrid sliding window (1024) + global attention (every 6th layer)" },
     { id: "gemma3-4b", family: "Gemma", name: "Gemma 3 4B", params: { n_layers: 34, kv_cache_layer_count: null, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 4000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "Hybrid sliding window (1024) + global attention (every 6th layer)" },
@@ -648,48 +669,97 @@
     { id: "mistral-small-3.1-24b", family: "Mistral", name: "Mistral Small 3.1 24B", params: { n_layers: 40, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 24000000000, weight_dtype_bytes: 2, block_size: 16 }, notes: "GQA, multimodal" },
   ];
 
+  // ═══════════════════════════════════════════════════════════════════
+  // Model Registry (single source of truth for model specs)
+  // ═══════════════════════════════════════════════════════════════════
+  const MODELS = {
+    "qwen3-27b": {
+      name: "Qwen3-27B",
+      family: "Qwen",
+      model_profile: { n_layers: 64, kv_cache_layer_count: 16, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 27781419504, weight_dtype_bytes: 2, block_size: 16 },
+      description: "Hybrid attention model with MLA (Multi-Latent Attention)"
+    },
+    "llama3.1-70b": {
+      name: "Llama 3.1 70B",
+      family: "Llama",
+      model_profile: { n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 70600000000, weight_dtype_bytes: 2, block_size: 16 },
+      description: "Meta's 70B parameter dense transformer"
+    },
+    "qwen2.5-72b": {
+      name: "Qwen2.5-72B",
+      family: "Qwen",
+      model_profile: { n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 72700000000, weight_dtype_bytes: 2, block_size: 16 },
+      description: "Alibaba's 72B parameter dense model"
+    },
+    "mistral-7b": {
+      name: "Mistral 7B",
+      family: "Mistral",
+      model_profile: { n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 7200000000, weight_dtype_bytes: 2, block_size: 16 },
+      description: "Mistral AI's efficient 7B model"
+    },
+    "llama3.1-8b": {
+      name: "Llama 3.1 8B",
+      family: "Llama",
+      model_profile: { n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 8000000000, weight_dtype_bytes: 2, block_size: 16 },
+      description: "Meta's 8B parameter efficient model"
+    },
+    "deepseek-v3-671b": {
+      name: "DeepSeek-V3-671B",
+      family: "DeepSeek",
+      model_profile: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, block_size: 16 },
+      description: "DeepSeek's 671B MoE model with MLA"
+    }
+  };
+
   // ═══════════════════════════════════════════════════════════════════
   // Deployment Preset Configurations
   // ═══════════════════════════════════════════════════════════════════
   const PRESETS = {
     "qwen3-27b-8xh20": {
-      model_profile: { n_layers: 64, kv_cache_layer_count: 16, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 27781419504, weight_dtype_bytes: 2, tp_size: 8, pp_size: 1, block_size: 16 },
+      model_id: "qwen3-27b",
+      model_profile: { ...MODELS["qwen3-27b"].model_profile, tp_size: 8, pp_size: 1 },
       prefill_savings_alpha: 0.8,
       heuristic_multi_agent: { concurrent_agents: 2048, shared_prefix_tokens: 4096, avg_new_tokens_per_turn: 4096, avg_turns_per_session: 8, private_window_tokens: 65536, curve_mode: "zipf_harmonic", zipf_s: 1.3, zipf_population_blocks: 4096, policy_efficiency: { strict_prefix_upper_bound: 1.0, lru_like: 0.6 } },
       deployment: { label: "1x8-h20", accelerator_count: 8, cards_per_machine: 8, machine_spec: "h20", gpu_memory_gb_per_card: 96, total_tps: 8.0, total_tps_unit: "cluster_total", baseline_per_card_tps: 1.0, planning_target_total_tps: 8.0, extra_capacity_tiers: [{ label: "HBM+1T per Machine", kv_gb_per_machine: 1024 }, { label: "HBM+10T per Machine", kv_gb_per_machine: 10240 }] }
     },
     "qwen3-27b-1xh20": {
-      model_profile: { n_layers: 64, kv_cache_layer_count: 16, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 27781419504, weight_dtype_bytes: 2, tp_size: 1, pp_size: 1, block_size: 16 },
+      model_id: "qwen3-27b",
+      model_profile: { ...MODELS["qwen3-27b"].model_profile, tp_size: 1, pp_size: 1 },
       prefill_savings_alpha: 0.8,
       heuristic_multi_agent: { concurrent_agents: 2048, shared_prefix_tokens: 4096, avg_new_tokens_per_turn: 4096, avg_turns_per_session: 8, private_window_tokens: 65536, curve_mode: "zipf_harmonic", zipf_s: 1.3, zipf_population_blocks: 4096, policy_efficiency: { strict_prefix_upper_bound: 1.0, lru_like: 0.6 } },
       deployment: { label: "1x1-h20", accelerator_count: 1, cards_per_machine: 1, machine_spec: "h20", gpu_memory_gb_per_card: 96, total_tps: 1.0, total_tps_unit: "cluster_total", baseline_per_card_tps: 1.0, planning_target_total_tps: 8.0, extra_capacity_tiers: [{ label: "HBM+1T per Machine", kv_gb_per_machine: 1024 }, { label: "HBM+10T per Machine", kv_gb_per_machine: 10240 }] }
     },
     "llama3.1-70b-8xa100": {
-      model_profile: { n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 70600000000, weight_dtype_bytes: 2, tp_size: 8, pp_size: 1, block_size: 16 },
+      model_id: "llama3.1-70b",
+      model_profile: { ...MODELS["llama3.1-70b"].model_profile, tp_size: 8, pp_size: 1 },
       prefill_savings_alpha: 0.8,
       heuristic_multi_agent: { concurrent_agents: 2048, shared_prefix_tokens: 4096, avg_new_tokens_per_turn: 4096, avg_turns_per_session: 8, private_window_tokens: 65536, curve_mode: "zipf_harmonic", zipf_s: 1.3, zipf_population_blocks: 4096, policy_efficiency: { strict_prefix_upper_bound: 1.0, lru_like: 0.6 } },
       deployment: { label: "1x8-a100", accelerator_count: 8, cards_per_machine: 8, machine_spec: "a100-80", gpu_memory_gb_per_card: 80, total_tps: 6.4, total_tps_unit: "cluster_total", baseline_per_card_tps: 0.8, planning_target_total_tps: 12.8, extra_capacity_tiers: [{ label: "HBM+1T per Machine", kv_gb_per_machine: 1024 }] }
     },
     "qwen2.5-72b-8xh100": {
-      model_profile: { n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 72700000000, weight_dtype_bytes: 2, tp_size: 8, pp_size: 1, block_size: 16 },
+      model_id: "qwen2.5-72b",
+      model_profile: { ...MODELS["qwen2.5-72b"].model_profile, tp_size: 8, pp_size: 1 },
       prefill_savings_alpha: 0.8,
       heuristic_multi_agent: { concurrent_agents: 2048, shared_prefix_tokens: 4096, avg_new_tokens_per_turn: 4096, avg_turns_per_session: 8, private_window_tokens: 65536, curve_mode: "zipf_harmonic", zipf_s: 1.3, zipf_population_blocks: 4096, policy_efficiency: { strict_prefix_upper_bound: 1.0, lru_like: 0.6 } },
       deployment: { label: "1x8-h100", accelerator_count: 8, cards_per_machine: 8, machine_spec: "h100", gpu_memory_gb_per_card: 80, total_tps: 12.0, total_tps_unit: "cluster_total", baseline_per_card_tps: 1.5, planning_target_total_tps: 24.0, extra_capacity_tiers: [{ label: "HBM+1T per Machine", kv_gb_per_machine: 1024 }] }
     },
     "mistral-7b-1xl40s": {
-      model_profile: { n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 7200000000, weight_dtype_bytes: 2, tp_size: 1, pp_size: 1, block_size: 16 },
+      model_id: "mistral-7b",
+      model_profile: { ...MODELS["mistral-7b"].model_profile, tp_size: 1, pp_size: 1 },
       prefill_savings_alpha: 0.8,
       heuristic_multi_agent: { concurrent_agents: 1024, shared_prefix_tokens: 4096, avg_new_tokens_per_turn: 2048, avg_turns_per_session: 4, private_window_tokens: 16384, curve_mode: "zipf_harmonic", zipf_s: 1.3, zipf_population_blocks: 4096, policy_efficiency: { strict_prefix_upper_bound: 1.0, lru_like: 0.6 } },
       deployment: { label: "1x1-l40s", accelerator_count: 1, cards_per_machine: 1, machine_spec: "l40s", gpu_memory_gb_per_card: 48, total_tps: 0.5, total_tps_unit: "cluster_total", baseline_per_card_tps: 0.5, planning_target_total_tps: 2.0, extra_capacity_tiers: [{ label: "HBM+512G", kv_gb_per_machine: 512 }] }
     },
     "llama3.1-8b-1xa100": {
-      model_profile: { n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 8000000000, weight_dtype_bytes: 2, tp_size: 1, pp_size: 1, block_size: 16 },
+      model_id: "llama3.1-8b",
+      model_profile: { ...MODELS["llama3.1-8b"].model_profile, tp_size: 1, pp_size: 1 },
       prefill_savings_alpha: 0.8,
       heuristic_multi_agent: { concurrent_agents: 2048, shared_prefix_tokens: 4096, avg_new_tokens_per_turn: 4096, avg_turns_per_session: 8, private_window_tokens: 65536, curve_mode: "zipf_harmonic", zipf_s: 1.3, zipf_population_blocks: 4096, policy_efficiency: { strict_prefix_upper_bound: 1.0, lru_like: 0.6 } },
       deployment: { label: "1x1-a100", accelerator_count: 1, cards_per_machine: 1, machine_spec: "a100-80", gpu_memory_gb_per_card: 80, total_tps: 0.8, total_tps_unit: "cluster_total", baseline_per_card_tps: 0.8, planning_target_total_tps: 6.4, extra_capacity_tiers: [{ label: "HBM+1T per Machine", kv_gb_per_machine: 1024 }] }
     },
     "deepseek-v3-671b-8xh200": {
-      model_profile: { n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, tp_size: 8, pp_size: 1, block_size: 16 },
+      model_id: "deepseek-v3-671b",
+      model_profile: { ...MODELS["deepseek-v3-671b"].model_profile, tp_size: 8, pp_size: 1 },
       prefill_savings_alpha: 0.8,
       heuristic_multi_agent: { concurrent_agents: 2048, shared_prefix_tokens: 4096, avg_new_tokens_per_turn: 4096, avg_turns_per_session: 8, private_window_tokens: 65536, curve_mode: "zipf_harmonic", zipf_s: 1.3, zipf_population_blocks: 4096, policy_efficiency: { strict_prefix_upper_bound: 1.0, lru_like: 0.6 } },
       deployment: { label: "1x8-h200", accelerator_count: 8, cards_per_machine: 8, machine_spec: "h200", gpu_memory_gb_per_card: 141, total_tps: 16.0, total_tps_unit: "cluster_total", baseline_per_card_tps: 2.0, planning_target_total_tps: 32.0, extra_capacity_tiers: [{ label: "HBM+2T per Machine", kv_gb_per_machine: 2048 }] }
@@ -906,6 +976,27 @@
       if (efficiency <= 0) return Infinity;
       return this.shared_prefix_tokens + this.totalPrivateWorkingSetTokens() / efficiency;
     }
+    perTurnReusablePrivateTokens() {
+      const result = [];
+      for (let step = 0; step < this.avg_turns_per_session; step++) {
+        result.push(Math.min(this.private_window_tokens, step * this.avg_new_tokens_per_turn));
+      }
+      return result;
+    }
+    perTurnContentHitRates() {
+      const rates = [];
+      for (let step = 0; step < this.avg_turns_per_session; step++) {
+        const privateReuse = Math.min(this.private_window_tokens, step * this.avg_new_tokens_per_turn);
+        const requestTokens = this.shared_prefix_tokens + this.avg_new_tokens_per_turn + privateReuse;
+        if (requestTokens <= 0) {
+          rates.push(0.0);
+        } else {
+          const hit = this.shared_prefix_tokens + privateReuse;
+          rates.push(Math.min(1.0, hit / requestTokens));
+        }
+      }
+      return rates;
+    }
   }
 
   function hitRateForCapacityTokens(capacityTokens, heuristic, efficiency) {
@@ -1064,6 +1155,7 @@
   // ═══════════════════════════════════════════════════════════════════
   let hitRateChart = null;
   let tpsMachinesChart = null;
+  let perTurnChart = null;
   let debounceTimer = null;
   let scenarioA = null;
   let modelPresetValues = null;
@@ -1121,7 +1213,111 @@
     };
   }
 
-  function loadPreset(name) {
+  // ═══════════════════════════════════════════════════════════════════
+  // Model Ranking Functions
+  // ═══════════════════════════════════════════════════════════════════
+  function formatBytes(bytes, decimals = 2) {
+    if (bytes === 0) return '0 B';
+    const k = 1024;
+    const dm = decimals < 0 ? 0 : decimals;
+    const sizes = ['B', 'KB', 'MB', 'GB', 'TB'];
+    const i = Math.floor(Math.log(bytes) / Math.log(k));
+    return parseFloat((bytes / Math.pow(k, i)).toFixed(dm)) + ' ' + sizes[i];
+  }
+
+  function calculateModelRanking(sequenceLength) {
+    const rankings = [];
+    for (const [modelId, modelData] of Object.entries(MODELS)) {
+      const model = modelData.model_profile;
+      const kvPerToken = kvBytesPerToken(model);
+      const kvPerTokenPerRank = kvBytesPerTokenPerRank(model);
+      const totalKvBytes = kvPerToken * sequenceLength;
+      const totalKvBytesPerRank = kvPerTokenPerRank * sequenceLength;
+      const totalKvGb = totalKvBytes / BYTES_PER_GB;
+      const totalKvGbPerRank = totalKvBytesPerRank / BYTES_PER_GB;
+
+      rankings.push({
+        modelId,
+        name: modelData.name,
+        family: modelData.family,
+        description: modelData.description,
+        kv_bytes_per_token: kvPerToken,
+        kv_bytes_per_token_per_rank: kvPerTokenPerRank,
+        total_kv_bytes: totalKvBytes,
+        total_kv_bytes_per_rank: totalKvBytesPerRank,
+        total_kv_gb: totalKvGb,
+        total_kv_gb_per_rank: totalKvGbPerRank,
+        layers: model.kv_cache_layer_count != null ? model.kv_cache_layer_count : model.n_layers,
+        n_kv_heads: model.n_kv_heads,
+        head_dim: model.head_dim
+      });
+    }
+    rankings.sort((a, b) => b.total_kv_gb - a.total_kv_gb);
+    return rankings;
+  }
+
+  let rankingSortState = { column: 'total_kv_gb', ascending: false };
+
+  function renderRankingTable(rankingData, sequenceLength, selectedModelId = null) {
+    const sortCol = rankingSortState.column;
+    const ascending = rankingSortState.ascending;
+    const sorted = [...rankingData].sort((a, b) => {
+      let va = a[sortCol], vb = b[sortCol];
+      if (typeof va === 'string') {
+        return ascending ? va.localeCompare(vb) : vb.localeCompare(va);
+      }
+      return ascending ? va - vb : vb - va;
+    });
+
+    const sortIcon = (col) => {
+      if (sortCol !== col) return '<span style="opacity:0.3">⇅</span>';
+      return ascending ? '↑' : '↓';
+    };
+
+    const seqLabel = sequenceLength >= 1024 ? `${(sequenceLength / 1024).toFixed(sequenceLength % 1024 === 0 ? 0 : 1)}K` : sequenceLength;
+
+    let html = `
+      <table class="ranking-table" style="width:100%; border-collapse:collapse; font-size:0.85rem; margin-top:12px;">
+        <thead>
+          <tr style="border-bottom:2px solid var(--border);">
+            <th style="padding:8px 6px; text-align:left; cursor:pointer;" data-sort="name">Model ${sortIcon('name')}</th>
+            <th style="padding:8px 6px; text-align:right; cursor:pointer;" data-sort="kv_bytes_per_token">KV/Token ${sortIcon('kv_bytes_per_token')}</th>
+            <th style="padding:8px 6px; text-align:right; cursor:pointer;" data-sort="kv_bytes_per_token_per_rank">KV/Token/Rank ${sortIcon('kv_bytes_per_token_per_rank')}</th>
+            <th style="padding:8px 6px; text-align:right; cursor:pointer;" data-sort="total_kv_gb">Total (${seqLabel} tokens) ${sortIcon('total_kv_gb')}</th>
+            <th style="padding:8px 6px; text-align:right; cursor:pointer;" data-sort="layers">KV Layers ${sortIcon('layers')}</th>
+            <th style="padding:8px 6px; text-align:right; cursor:pointer;" data-sort="n_kv_heads">KV Heads ${sortIcon('n_kv_heads')}</th>
+            <th style="padding:8px 6px; text-align:right; cursor:pointer;" data-sort="head_dim">Head Dim ${sortIcon('head_dim')}</th>
+          </tr>
+        </thead>
+        <tbody>`;
+
+    sorted.forEach((item, idx) => {
+      const isSelected = item.modelId === selectedModelId;
+      const rowStyle = isSelected
+        ? 'background:rgba(74,108,247,0.1); font-weight:600;'
+        : (idx % 2 === 0 ? '' : 'background:rgba(0,0,0,0.02);');
+      html += `
+          <tr style="${rowStyle} border-bottom:1px solid var(--border);">
+            <td style="padding:8px 6px;">
+              <div style="font-weight:600;">${item.name}</div>
+              <div style="font-size:0.75rem; color:var(--text-muted);">${item.family}</div>
+            </td>
+            <td style="padding:8px 6px; text-align:right;">${formatBytes(item.kv_bytes_per_token)}</td>
+            <td style="padding:8px 6px; text-align:right;">${formatBytes(item.kv_bytes_per_token_per_rank)}</td>
+            <td style="padding:8px 6px; text-align:right;">${item.total_kv_gb.toFixed(2)} GB</td>
+            <td style="padding:8px 6px; text-align:right;">${item.layers}</td>
+            <td style="padding:8px 6px; text-align:right;">${item.n_kv_heads}</td>
+            <td style="padding:8px 6px; text-align:right;">${item.head_dim}</td>
+          </tr>`;
+    });
+
+    html += `
+        </tbody>
+      </table>`;
+    return html;
+  }
+
+
     const p = PRESETS[name];
     if (!p) return;
     const m = p.model_profile;
@@ -1180,6 +1376,7 @@
     updateCurveModeVisibility();
     checkModelModified();
     recalculate();
+    if (typeof updateRankingTable === 'function') updateRankingTable();
   }
 
   function loadDatasetPreset(datasetId) {
@@ -1239,6 +1436,7 @@
     updateCurveModeVisibility();
     checkModelModified();
     recalculate();
+    if (typeof updateRankingTable === 'function') updateRankingTable();
   }
 
   function findModelForPreset(modelProfile) {
@@ -1272,6 +1470,7 @@
     modelPresetValues = { ...p, tp_size: +document.getElementById("tp_size").value || 1, pp_size: +document.getElementById("pp_size").value || 1 };
     checkModelModified();
     scheduleRecalc();
+    if (typeof updateRankingTable === 'function') updateRankingTable();
   }
 
   function checkModelModified() {
@@ -1614,6 +1813,76 @@
     }).join("");
   }
 
+  function renderPerTurnChart(config) {
+    const heuristic = new MultiAgentHeuristic(config.heuristic_multi_agent);
+    const hitRates = heuristic.perTurnContentHitRates();
+    const privateTokens = heuristic.perTurnReusablePrivateTokens();
+    const labels = hitRates.map((_, i) => "Turn " + (i + 1));
+    const card = document.getElementById("per-turn-chart-card");
+    card.style.display = "";
+
+    const ctx = document.getElementById("perTurnChart").getContext("2d");
+    if (perTurnChart) perTurnChart.destroy();
+    perTurnChart = new Chart(ctx, {
+      type: "bar",
+      data: {
+        labels,
+        datasets: [
+          {
+            label: "Content Hit Rate (%)",
+            data: hitRates.map(r => r * 100),
+            backgroundColor: "rgba(88,166,255,0.5)",
+            borderColor: "#58a6ff",
+            borderWidth: 1,
+            yAxisID: "y",
+            order: 2
+          },
+          {
+            label: "Reusable Private Tokens",
+            data: privateTokens,
+            type: "line",
+            borderColor: "#3fb950",
+            backgroundColor: "rgba(63,185,80,0.1)",
+            tension: 0.3,
+            pointRadius: 4,
+            fill: false,
+            yAxisID: "y1",
+            order: 1
+          }
+        ]
+      },
+      options: {
+        responsive: true,
+        maintainAspectRatio: false,
+        scales: {
+          y: {
+            position: "left",
+            min: 0,
+            max: 100,
+            title: { display: true, text: "Content Hit Rate (%)", color: "#8b949e" },
+            grid: { color: "rgba(48,54,61,0.5)" },
+            ticks: { color: "#8b949e" }
+          },
+          y1: {
+            position: "right",
+            min: 0,
+            title: { display: true, text: "Reusable Private Tokens", color: "#8b949e" },
+            grid: { drawOnChartArea: false },
+            ticks: { color: "#8b949e" }
+          },
+          x: {
+            title: { display: true, text: "Conversation Turn", color: "#8b949e" },
+            grid: { color: "rgba(48,54,61,0.5)" },
+            ticks: { color: "#8b949e" }
+          }
+        },
+        plugins: {
+          legend: { labels: { color: "#c9d1d9" } }
+        }
+      }
+    });
+  }
+
   function renderComparison(resultB) {
     if (!scenarioA) return;
     const panel = document.getElementById("compare-panel");
@@ -1666,6 +1935,7 @@ <h4>Scenario B (Current)</h4>
       renderHitRateChart(result.rows, result.summary.contentHitRate, result.summary);
       renderTpsMachinesChart(config, result.summary);
       renderSensitivity(config);
+      renderPerTurnChart(config);
       if (scenarioA) renderComparison(result);
       validateInputs(config);
     } catch (e) {
@@ -2227,10 +2497,49 @@ <h2>Analysis Results</h2>
 
     // Wire all inputs for recalculation
     document.querySelectorAll("input, select").forEach(el => {
-      if (el.id === "preset-select" || el.id === "model-preset-select" || el.id === "model-search" || el.id === "gpu-hardware-select") return;
+      if (el.id === "preset-select" || el.id === "model-preset-select" || el.id === "model-search" || el.id === "gpu-hardware-select" || el.id === "ranking-sequence-length") return;
       el.addEventListener("input", scheduleRecalc);
       el.addEventListener("change", scheduleRecalc);
     });
+
+    // Model Ranking UI
+    function updateRankingTable() {
+      const seqLength = parseInt(document.getElementById("ranking-sequence-length").value) || 8192;
+      const rankings = calculateModelRanking(seqLength);
+
+      // Determine selected model from current preset
+      const presetSelect = document.getElementById("preset-select");
+      let selectedModelId = null;
+      if (presetSelect && presetSelect.value !== "custom" && !presetSelect.value.startsWith("dataset:")) {
+        const preset = PRESETS[presetSelect.value];
+        if (preset && preset.model_id) {
+          selectedModelId = preset.model_id;
+        }
+      }
+
+      const container = document.getElementById("ranking-table-container");
+      container.innerHTML = renderRankingTable(rankings, seqLength, selectedModelId);
+
+      // Wire up sort handlers
+      container.querySelectorAll("th[data-sort]").forEach(th => {
+        th.addEventListener("click", () => {
+          const col = th.dataset.sort;
+          if (rankingSortState.column === col) {
+            rankingSortState.ascending = !rankingSortState.ascending;
+          } else {
+            rankingSortState.column = col;
+            rankingSortState.ascending = col === 'name' || col === 'family';
+          }
+          updateRankingTable();
+        });
+      });
+    }
+
+    document.getElementById("ranking-sequence-length").addEventListener("input", updateRankingTable);
+    document.getElementById("ranking-sequence-length").addEventListener("change", updateRankingTable);
+
+    // Initial ranking calculation
+    updateRankingTable();
   });
   </script>
 </body>
diff --git a/website/ranking.html b/website/ranking.html
new file mode 100644
index 0000000..6fdfd9b
--- /dev/null
+++ b/website/ranking.html
@@ -0,0 +1,584 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+  <title>KVCache Model Ranking</title>
+  <style>
+    *, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; }
+
+    :root {
+      --bg: #0d1117;
+      --surface: #161b22;
+      --surface2: #21262d;
+      --border: #30363d;
+      --text: #c9d1d9;
+      --text-muted: #8b949e;
+      --accent: #58a6ff;
+      --radius: 8px;
+    }
+
+    html { font-size: 15px; }
+    body {
+      background: var(--bg);
+      color: var(--text);
+      font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
+      line-height: 1.6;
+      min-height: 100vh;
+      padding: 32px;
+    }
+
+    header {
+      max-width: 1400px;
+      margin: 0 auto 32px;
+    }
+
+    h1 {
+      font-size: 1.8rem;
+      font-weight: 600;
+      margin-bottom: 8px;
+    }
+
+    .subtitle {
+      color: var(--text-muted);
+      font-size: 0.95rem;
+    }
+
+    .controls {
+      max-width: 1400px;
+      margin: 0 auto 24px;
+      display: flex;
+      gap: 12px;
+      align-items: center;
+    }
+
+    .controls label {
+      color: var(--text-muted);
+      font-size: 0.9rem;
+    }
+
+    .preset-buttons {
+      display: flex;
+      gap: 8px;
+      flex-wrap: wrap;
+    }
+
+    .preset-btn {
+      background: var(--surface);
+      border: 1px solid var(--border);
+      color: var(--text);
+      padding: 8px 16px;
+      border-radius: var(--radius);
+      cursor: pointer;
+      font-size: 0.9rem;
+      transition: all 0.15s;
+    }
+
+    .preset-btn:hover {
+      background: var(--surface2);
+      border-color: var(--accent);
+    }
+
+    .preset-btn.active {
+      background: var(--accent);
+      border-color: var(--accent);
+      color: #fff;
+    }
+
+    .input-group {
+      display: flex;
+      gap: 16px;
+      align-items: center;
+    }
+
+    .input-field {
+      display: flex;
+      align-items: center;
+      gap: 8px;
+    }
+
+    .input-field input {
+      background: var(--surface);
+      border: 1px solid var(--border);
+      color: var(--text);
+      padding: 8px 12px;
+      border-radius: var(--radius);
+      font-size: 0.9rem;
+      width: 80px;
+      transition: all 0.15s;
+    }
+
+    .input-field input:hover {
+      border-color: var(--accent);
+    }
+
+    .input-field input:focus {
+      outline: none;
+      border-color: var(--accent);
+      background: var(--surface2);
+    }
+
+    #family-filter {
+      background: var(--surface);
+      border: 1px solid var(--border);
+      color: var(--text);
+      padding: 8px 12px;
+      border-radius: var(--radius);
+      font-size: 0.9rem;
+      cursor: pointer;
+      transition: all 0.15s;
+    }
+
+    #family-filter:hover, #family-filter:focus {
+      border-color: var(--accent);
+      outline: none;
+    }
+
+    .table-container {
+      max-width: 1400px;
+      margin: 0 auto;
+      background: var(--surface);
+      border: 1px solid var(--border);
+      border-radius: var(--radius);
+      overflow: hidden;
+    }
+
+    table {
+      width: 100%;
+      border-collapse: collapse;
+    }
+
+    thead {
+      background: var(--surface2);
+      border-bottom: 2px solid var(--border);
+    }
+
+    th {
+      padding: 12px 16px;
+      text-align: left;
+      font-weight: 600;
+      font-size: 0.85rem;
+      color: var(--text);
+      cursor: pointer;
+      user-select: none;
+      white-space: nowrap;
+    }
+
+    th:hover {
+      background: rgba(88, 166, 255, 0.1);
+    }
+
+    th .sort-icon {
+      margin-left: 4px;
+      opacity: 0.4;
+    }
+
+    th.sorted .sort-icon {
+      opacity: 1;
+      color: var(--accent);
+    }
+
+    tbody tr {
+      border-bottom: 1px solid var(--border);
+      transition: background 0.1s;
+    }
+
+    tbody tr:hover {
+      background: var(--surface2);
+    }
+
+    tbody tr:last-child {
+      border-bottom: none;
+    }
+
+    td {
+      padding: 12px 16px;
+      font-size: 0.9rem;
+    }
+
+    .model-name {
+      font-weight: 600;
+      color: var(--text);
+    }
+
+    .model-family {
+      font-size: 0.8rem;
+      color: var(--text-muted);
+    }
+
+    .num-cell {
+      text-align: right;
+      font-variant-numeric: tabular-nums;
+    }
+
+    .highlight {
+      color: var(--accent);
+      font-weight: 600;
+    }
+
+    footer {
+      max-width: 1400px;
+      margin: 32px auto 0;
+      text-align: center;
+      color: var(--text-muted);
+      font-size: 0.85rem;
+    }
+
+    footer a {
+      color: var(--accent);
+      text-decoration: none;
+    }
+
+    footer a:hover {
+      text-decoration: underline;
+    }
+  </style>
+</head>
+<body>
+  <header>
+    <h1>KVCache Model Ranking</h1>
+    <p class="subtitle">Compare KV cache size requirements across popular LLM models</p>
+  </header>
+
+  <div class="controls">
+    <label>Sequence Length:</label>
+    <div class="preset-buttons" id="preset-buttons"></div>
+  </div>
+
+  <div class="controls">
+    <label>Parallelism:</label>
+    <div class="input-group">
+      <div class="input-field">
+        <label for="tp-input">TP</label>
+        <input type="number" id="tp-input" value="1" min="1" max="16" />
+      </div>
+      <div class="input-field">
+        <label for="pp-input">PP</label>
+        <input type="number" id="pp-input" value="1" min="1" max="16" />
+      </div>
+    </div>
+  </div>
+
+  <div class="controls">
+    <label>Filter by family:</label>
+    <select id="family-filter">
+      <option value="all">All Families</option>
+    </select>
+  </div>
+
+  <div class="table-container">
+    <table id="ranking-table">
+      <thead>
+        <tr>
+          <th data-column="name">Model <span class="sort-icon">↕</span></th>
+          <th data-column="layers" data-type="number">KV Layers <span class="sort-icon">↕</span></th>
+          <th data-column="n_kv_heads" data-type="number">KV Heads <span class="sort-icon">↕</span></th>
+          <th data-column="head_dim" data-type="number">Head Dim <span class="sort-icon">↕</span></th>
+          <th data-column="tp_size" data-type="number">TP <span class="sort-icon">↕</span></th>
+          <th data-column="pp_size" data-type="number">PP <span class="sort-icon">↕</span></th>
+          <th data-column="parameters" data-type="number">Parameters <span class="sort-icon">↕</span></th>
+          <th data-column="weight_gb" data-type="number">Weight (GB) <span class="sort-icon">↕</span></th>
+          <th data-column="kv_per_token" data-type="number">KV/Token <span class="sort-icon">↕</span></th>
+          <th data-column="kv_per_token_rank" data-type="number">KV/Token/Rank <span class="sort-icon">↕</span></th>
+          <th data-column="total_kv" data-type="number" class="highlight">Total KV <span class="sort-icon">↕</span></th>
+          <th data-column="total_kv_rank" data-type="number" class="highlight">Total KV/Rank <span class="sort-icon">↕</span></th>
+        </tr>
+      </thead>
+      <tbody id="ranking-tbody"></tbody>
+    </table>
+  </div>
+
+  <footer>
+    <p>KVCache Upper Bound Oracle — <a href="index.html">Dashboard</a> | <a href="calculator.html">Calculator</a></p>
+  </footer>
+
+  <script>
+    const MODELS = {
+      // === DeepSeek ===
+      "deepseek-v4-pro": { name: "DeepSeek V4 Pro", family: "DeepSeek", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, notes: "V4 Hybrid (compress_ratios, sw=128)" },
+      "deepseek-v4-flash": { name: "DeepSeek V4 Flash", family: "DeepSeek", n_layers: 43, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, notes: "V4 Hybrid (compress_ratios, sw=128)" },
+      "deepseek-v3.2": { name: "DeepSeek V3.2", family: "DeepSeek", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 685000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "deepseek-v3": { name: "DeepSeek V3", family: "DeepSeek", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "deepseek-r1": { name: "DeepSeek R1", family: "DeepSeek", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 671000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "deepseek-v2-236b": { name: "DeepSeek-V2 236B", family: "DeepSeek", n_layers: 60, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 236000000000, weight_dtype_bytes: 2, notes: "MLA + MoE (compressed KV latent: kv_lora_rank=512)" },
+      "deepseek-v3.1-685b": { name: "DeepSeek-V3.1 685B", family: "DeepSeek", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 685000000000, weight_dtype_bytes: 2, notes: "MLA + MoE 256E/8A (compressed KV latent: kv_lora_rank=512)" },
+
+      // === GLM ===
+      "glm-5": { name: "GLM-5", family: "GLM", n_layers: 78, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 756000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "glm-5.1": { name: "GLM-5.1", family: "GLM", n_layers: 78, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 756000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "glm-5.2": { name: "GLM-5.2", family: "GLM", n_layers: 78, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 756000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "glm4-9b": { name: "GLM-4 9B", family: "GLM", n_layers: 40, kv_cache_layer_count: null, n_kv_heads: 2, head_dim: 128, dtype_bytes: 2, parameter_count: 9400000000, weight_dtype_bytes: 2, notes: "MQA" },
+      "glm-z1-32b": { name: "GLM-Z1 32B", family: "GLM", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 2, head_dim: 128, dtype_bytes: 2, parameter_count: 32000000000, weight_dtype_bytes: 2, notes: "GQA 24:1 ratio" },
+
+      // === Kimi ===
+      "kimi-k2.5": { name: "Kimi K2.5", family: "Kimi", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 1000000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "kimi-k2.6": { name: "Kimi K2.6", family: "Kimi", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 1000000000000, weight_dtype_bytes: 2, notes: "MLA, kv_lora_rank=512" },
+      "kimi-k2-1t": { name: "Kimi K2 1T", family: "Kimi", n_layers: 61, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 512, dtype_bytes: 2, parameter_count: 1000000000000, weight_dtype_bytes: 2, notes: "MLA + MoE 384E/8A (DeepSeek-V3 arch, compressed KV latent: kv_lora_rank=512)" },
+
+      // === Qwen3.6 (Hybrid linear+full) ===
+      "qwen3.6-27b": { name: "Qwen3.6-27B", family: "Qwen3.6", n_layers: 64, kv_cache_layer_count: 16, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 27000000000, weight_dtype_bytes: 2, notes: "Hybrid: 16 full + 48 linear attn layers" },
+      "qwen3.6-35b-a3b": { name: "Qwen3.6-35B-A3B", family: "Qwen3.6", n_layers: 40, kv_cache_layer_count: 10, n_kv_heads: 2, head_dim: 256, dtype_bytes: 2, parameter_count: 35000000000, weight_dtype_bytes: 2, notes: "Hybrid: 10 full + 30 linear attn layers" },
+
+      // === Qwen3.5 (Hybrid linear+full) ===
+      "qwen3.5-397b-a17b": { name: "Qwen3.5-397B-A17B", family: "Qwen3.5", n_layers: 60, kv_cache_layer_count: 15, n_kv_heads: 2, head_dim: 256, dtype_bytes: 2, parameter_count: 397000000000, weight_dtype_bytes: 2, notes: "Hybrid: 15 full + 45 linear attn layers" },
+      "qwen3.5-122b-a10b": { name: "Qwen3.5-122B-A10B", family: "Qwen3.5", n_layers: 48, kv_cache_layer_count: 12, n_kv_heads: 2, head_dim: 256, dtype_bytes: 2, parameter_count: 122000000000, weight_dtype_bytes: 2, notes: "Hybrid: 12 full + 36 linear attn layers" },
+      "qwen3.5-35b-a3b": { name: "Qwen3.5-35B-A3B", family: "Qwen3.5", n_layers: 40, kv_cache_layer_count: 10, n_kv_heads: 2, head_dim: 256, dtype_bytes: 2, parameter_count: 35000000000, weight_dtype_bytes: 2, notes: "Hybrid: 10 full + 30 linear attn layers" },
+      "qwen3.5-27b": { name: "Qwen3.5-27B", family: "Qwen3.5", n_layers: 64, kv_cache_layer_count: 16, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 27000000000, weight_dtype_bytes: 2, notes: "Hybrid: 16 full + 48 linear attn layers" },
+      "qwen3.5-9b": { name: "Qwen3.5-9B", family: "Qwen3.5", n_layers: 32, kv_cache_layer_count: 8, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 9000000000, weight_dtype_bytes: 2, notes: "Hybrid: 8 full + 24 linear attn layers" },
+      "qwen3.5-4b": { name: "Qwen3.5-4B", family: "Qwen3.5", n_layers: 32, kv_cache_layer_count: 8, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 4000000000, weight_dtype_bytes: 2, notes: "Hybrid: 8 full + 24 linear attn layers" },
+      "qwen3.5-2b": { name: "Qwen3.5-2B", family: "Qwen3.5", n_layers: 24, kv_cache_layer_count: 6, n_kv_heads: 2, head_dim: 256, dtype_bytes: 2, parameter_count: 2000000000, weight_dtype_bytes: 2, notes: "Hybrid: 6 full + 18 linear attn layers" },
+      "qwen3.5-0.8b": { name: "Qwen3.5-0.8B", family: "Qwen3.5", n_layers: 24, kv_cache_layer_count: 6, n_kv_heads: 2, head_dim: 256, dtype_bytes: 2, parameter_count: 800000000, weight_dtype_bytes: 2, notes: "Hybrid: 6 full + 18 linear attn layers" },
+
+      // === Qwen3 (GQA) ===
+      "qwen3-235b-a22b": { name: "Qwen3-235B-A22B", family: "Qwen3", n_layers: 94, kv_cache_layer_count: null, n_kv_heads: 4, head_dim: 128, dtype_bytes: 2, parameter_count: 235000000000, weight_dtype_bytes: 2, notes: "GQA + MoE 128E/8A" },
+      "qwen3-32b": { name: "Qwen3-32B", family: "Qwen3", n_layers: 64, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 32000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen3-30b-a3b": { name: "Qwen3-30B-A3B", family: "Qwen3", n_layers: 48, kv_cache_layer_count: null, n_kv_heads: 4, head_dim: 128, dtype_bytes: 2, parameter_count: 30000000000, weight_dtype_bytes: 2, notes: "GQA + MoE 128E/8A" },
+      "qwen3-14b": { name: "Qwen3-14B", family: "Qwen3", n_layers: 40, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 14000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen3-8b": { name: "Qwen3-8B", family: "Qwen3", n_layers: 36, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 8000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen3-4b": { name: "Qwen3-4B", family: "Qwen3", n_layers: 36, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 4000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen3-1.7b": { name: "Qwen3-1.7B", family: "Qwen3", n_layers: 28, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 1700000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen3-0.6b": { name: "Qwen3-0.6B", family: "Qwen3", n_layers: 28, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 600000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen3-27b": { name: "Qwen3-27B (MLA)", family: "Qwen3", n_layers: 64, kv_cache_layer_count: 16, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 27781419504, weight_dtype_bytes: 2, notes: "MLA — only 16 KV layers" },
+
+      // === Qwen2.5 (GQA) ===
+      "qwen2.5-72b": { name: "Qwen2.5-72B", family: "Qwen2.5", n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 72700000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen2.5-32b": { name: "Qwen2.5-32B", family: "Qwen2.5", n_layers: 64, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 32000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen2.5-14b": { name: "Qwen2.5-14B", family: "Qwen2.5", n_layers: 48, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 14000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen2.5-7b": { name: "Qwen2.5-7B", family: "Qwen2.5", n_layers: 28, kv_cache_layer_count: null, n_kv_heads: 4, head_dim: 128, dtype_bytes: 2, parameter_count: 7600000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "qwen2.5-coder-32b": { name: "Qwen2.5-Coder-32B", family: "Qwen2.5", n_layers: 64, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 32000000000, weight_dtype_bytes: 2, notes: "GQA" },
+
+      // === Llama (GQA) ===
+      "llama3.1-8b": { name: "Llama 3.1 8B", family: "Llama", n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 8000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "llama3.1-70b": { name: "Llama 3.1 70B", family: "Llama", n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 70600000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "llama3.3-70b": { name: "Llama 3.3 70B", family: "Llama", n_layers: 80, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 70600000000, weight_dtype_bytes: 2, notes: "GQA" },
+
+      // === Gemma (Hybrid) ===
+      "gemma-4-e2b": { name: "Gemma 4 E2B", family: "Gemma", n_layers: 35, kv_cache_layer_count: null, n_kv_heads: 1, head_dim: 256, dtype_bytes: 2, parameter_count: 2000000000, weight_dtype_bytes: 2, notes: "Hybrid: 3 full + sliding window (sw=512)" },
+      "gemma-4-e4b": { name: "Gemma 4 E4B", family: "Gemma", n_layers: 42, kv_cache_layer_count: null, n_kv_heads: 2, head_dim: 256, dtype_bytes: 2, parameter_count: 4000000000, weight_dtype_bytes: 2, notes: "Hybrid: 4 full + sliding window (sw=512)" },
+      "gemma-4-26b-a4b": { name: "Gemma 4 26B-A4B", family: "Gemma", n_layers: 30, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 256, dtype_bytes: 2, parameter_count: 26000000000, weight_dtype_bytes: 2, notes: "Hybrid: 5 full + sliding window (sw=1024)" },
+      "gemma-4-31b": { name: "Gemma 4 31B", family: "Gemma", n_layers: 60, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 256, dtype_bytes: 2, parameter_count: 31000000000, weight_dtype_bytes: 2, notes: "Hybrid: 10 full + sliding window (sw=1024)" },
+      "gemma3-27b": { name: "Gemma 3 27B", family: "Gemma", n_layers: 62, kv_cache_layer_count: null, n_kv_heads: 16, head_dim: 128, dtype_bytes: 2, parameter_count: 27000000000, weight_dtype_bytes: 2, notes: "Hybrid sliding window (1024) + global (every 6th)" },
+      "gemma3-12b": { name: "Gemma 3 12B", family: "Gemma", n_layers: 48, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 256, dtype_bytes: 2, parameter_count: 12000000000, weight_dtype_bytes: 2, notes: "Hybrid sliding window (1024) + global (every 6th)" },
+      "gemma3-4b": { name: "Gemma 3 4B", family: "Gemma", n_layers: 34, kv_cache_layer_count: null, n_kv_heads: 4, head_dim: 256, dtype_bytes: 2, parameter_count: 4000000000, weight_dtype_bytes: 2, notes: "Hybrid sliding window (1024) + global (every 6th)" },
+
+      // === Cohere ===
+      "cohere-command-r-v01": { name: "Cohere Command R v01", family: "Cohere", n_layers: 40, kv_cache_layer_count: null, n_kv_heads: 64, head_dim: 128, dtype_bytes: 2, parameter_count: 35000000000, weight_dtype_bytes: 2, notes: "GQA (MHA)" },
+      "cohere-command-r-plus": { name: "Cohere Command R+", family: "Cohere", n_layers: 64, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 104000000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "cohere-command-r7b": { name: "Cohere Command R7B 12-2024", family: "Cohere", n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 7000000000, weight_dtype_bytes: 2, notes: "Hybrid: 8 full + sliding window (sw=4096)" },
+      "cohere-command-a": { name: "Cohere Command A 03-2025", family: "Cohere", n_layers: 64, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 111000000000, weight_dtype_bytes: 2, notes: "Hybrid: 16 full + sliding window (sw=4096)" },
+      "cohere-command-a-plus": { name: "Cohere Command A Plus 05-2026", family: "Cohere", n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 111000000000, weight_dtype_bytes: 2, notes: "Hybrid: 8 full + sliding window (sw=4096)" },
+
+      // === MiMo ===
+      "mimo-v2.5": { name: "MiMo-V2.5", family: "MiMo", n_layers: 48, kv_cache_layer_count: null, n_kv_heads: 4, head_dim: 192, dtype_bytes: 2, parameter_count: 7000000000, weight_dtype_bytes: 2, notes: "Hybrid: 9 full + sliding window (sw=128)" },
+      "mimo-v2.5-pro": { name: "MiMo-V2.5-Pro", family: "MiMo", n_layers: 70, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 192, dtype_bytes: 2, parameter_count: 14000000000, weight_dtype_bytes: 2, notes: "Hybrid: 10 full + sliding window (sw=128)" },
+
+      // === MiniMax ===
+      "minimax-m2": { name: "MiniMax M2", family: "MiniMax", n_layers: 62, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 230000000000, weight_dtype_bytes: 2, notes: "GQA + MoE 256E/8A" },
+      "minimax-m2.1": { name: "MiniMax M2.1", family: "MiniMax", n_layers: 62, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 230000000000, weight_dtype_bytes: 2, notes: "GQA + MoE 256E/8A" },
+      "minimax-m2.5": { name: "MiniMax M2.5", family: "MiniMax", n_layers: 62, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 230000000000, weight_dtype_bytes: 2, notes: "GQA + MoE 256E/8A" },
+      "minimax-m2.7": { name: "MiniMax M2.7", family: "MiniMax", n_layers: 62, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 230000000000, weight_dtype_bytes: 2, notes: "GQA + MoE 256E/8A" },
+
+      // === Mistral ===
+      "mistral-7b": { name: "Mistral 7B", family: "Mistral", n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 7200000000, weight_dtype_bytes: 2, notes: "GQA" },
+      "mistral-small-3.1-24b": { name: "Mistral Small 3.1 24B", family: "Mistral", n_layers: 40, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 24000000000, weight_dtype_bytes: 2, notes: "GQA, multimodal" },
+      "mixtral-8x22b": { name: "Mixtral 8x22B", family: "Mistral", n_layers: 56, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 141000000000, weight_dtype_bytes: 2, notes: "MoE, GQA" },
+
+      // === Yi ===
+      "yi1.5-34b": { name: "Yi-1.5 34B", family: "Yi", n_layers: 60, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 34400000000, weight_dtype_bytes: 2, notes: "GQA" },
+
+      // === Phi ===
+      "phi4-mini-3.8b": { name: "Phi-4-mini 3.8B", family: "Phi", n_layers: 32, kv_cache_layer_count: null, n_kv_heads: 8, head_dim: 128, dtype_bytes: 2, parameter_count: 3800000000, weight_dtype_bytes: 2, notes: "GQA + LongRoPE" },
+    };
+
+    const SEQUENCE_LENGTHS = [1024, 4096, 8192, 16384, 32768, 131072];
+    const BYTES_PER_GB = 1024 * 1024 * 1024;
+
+    let currentSequenceLength = 8192;
+    let currentTpSize = 1;
+    let currentPpSize = 1;
+    let currentFamily = 'all';
+    let sortColumn = 'total_kv';
+    let sortDirection = 'desc';
+
+    function kvBytesPerToken(model) {
+      const layers = model.kv_cache_layer_count || model.n_layers;
+      return 2 * layers * model.n_kv_heads * model.head_dim * model.dtype_bytes;
+    }
+
+    function kvBytesPerTokenPerRank(model) {
+      const shard = currentTpSize * currentPpSize;
+      return kvBytesPerToken(model) / shard;
+    }
+
+    function formatBytes(bytes) {
+      if (bytes < 1024) return bytes.toFixed(0) + ' B';
+      if (bytes < 1024 * 1024) return (bytes / 1024).toFixed(1) + ' KB';
+      if (bytes < 1024 * 1024 * 1024) return (bytes / (1024 * 1024)).toFixed(1) + ' MB';
+      return (bytes / BYTES_PER_GB).toFixed(2) + ' GB';
+    }
+
+    function formatParameters(count) {
+      if (count >= 1e9) return (count / 1e9).toFixed(1) + 'B';
+      if (count >= 1e6) return (count / 1e6).toFixed(1) + 'M';
+      return count.toLocaleString();
+    }
+
+    function computeRanking() {
+      return Object.entries(MODELS)
+        .filter(([id, model]) => currentFamily === 'all' || model.family === currentFamily)
+        .map(([id, model]) => {
+          const layers = model.kv_cache_layer_count || model.n_layers;
+          const kvPerToken = kvBytesPerToken(model);
+          const kvPerTokenRank = kvBytesPerTokenPerRank(model);
+          const totalKv = kvPerToken * currentSequenceLength;
+          const totalKvRank = kvPerTokenRank * currentSequenceLength;
+          const weightBytes = model.parameter_count ? model.parameter_count * model.weight_dtype_bytes : null;
+          const weightGb = weightBytes ? weightBytes / BYTES_PER_GB : null;
+
+          return {
+            id,
+            name: model.name,
+            family: model.family,
+            layers,
+            n_kv_heads: model.n_kv_heads,
+            head_dim: model.head_dim,
+            tp_size: currentTpSize,
+            pp_size: currentPpSize,
+            parameters: model.parameter_count,
+            weight_gb: weightGb,
+            kv_per_token: kvPerToken,
+            kv_per_token_rank: kvPerTokenRank,
+            total_kv: totalKv,
+            total_kv_rank: totalKvRank,
+            notes: model.notes || '',
+          };
+        });
+    }
+
+    function sortRanking(ranking) {
+      return ranking.sort((a, b) => {
+        let aVal = a[sortColumn];
+        let bVal = b[sortColumn];
+
+        if (typeof aVal === 'string') {
+          return sortDirection === 'asc'
+            ? aVal.localeCompare(bVal)
+            : bVal.localeCompare(aVal);
+        }
+
+        return sortDirection === 'asc' ? aVal - bVal : bVal - aVal;
+      });
+    }
+
+    function renderTable() {
+      const ranking = sortRanking(computeRanking());
+      const tbody = document.getElementById('ranking-tbody');
+      tbody.innerHTML = '';
+
+      ranking.forEach(model => {
+        const tr = document.createElement('tr');
+        tr.innerHTML = `
+          <td>
+            <div class="model-name">${model.name}</div>
+            <div class="model-family">${model.family}</div>
+          </td>
+          <td class="num-cell">${model.layers}</td>
+          <td class="num-cell">${model.n_kv_heads}</td>
+          <td class="num-cell">${model.head_dim}</td>
+          <td class="num-cell">${model.tp_size}</td>
+          <td class="num-cell">${model.pp_size}</td>
+          <td class="num-cell">${model.parameters ? formatParameters(model.parameters) : 'N/A'}</td>
+          <td class="num-cell">${model.weight_gb !== null ? model.weight_gb.toFixed(1) : 'N/A'}</td>
+          <td class="num-cell">${formatBytes(model.kv_per_token)}</td>
+          <td class="num-cell">${formatBytes(model.kv_per_token_rank)}</td>
+          <td class="num-cell highlight">${formatBytes(model.total_kv)}</td>
+          <td class="num-cell highlight">${formatBytes(model.total_kv_rank)}</td>
+        `;
+        tbody.appendChild(tr);
+      });
+
+      // Update sort indicators
+      document.querySelectorAll('th').forEach(th => {
+        th.classList.remove('sorted');
+        const icon = th.querySelector('.sort-icon');
+        if (icon) icon.textContent = '↕';
+      });
+
+      const activeTh = document.querySelector(`th[data-column="${sortColumn}"]`);
+      if (activeTh) {
+        activeTh.classList.add('sorted');
+        const icon = activeTh.querySelector('.sort-icon');
+        if (icon) icon.textContent = sortDirection === 'asc' ? '↑' : '↓';
+      }
+    }
+
+    function renderPresetButtons() {
+      const container = document.getElementById('preset-buttons');
+      SEQUENCE_LENGTHS.forEach(len => {
+        const btn = document.createElement('button');
+        btn.className = 'preset-btn' + (len === currentSequenceLength ? ' active' : '');
+        btn.textContent = len >= 1024 ? (len / 1024) + 'K' : len;
+        btn.onclick = () => {
+          currentSequenceLength = len;
+          renderPresetButtons();
+          renderTable();
+        };
+        container.appendChild(btn);
+      });
+    }
+
+    function setupSorting() {
+      document.querySelectorAll('th[data-column]').forEach(th => {
+        th.addEventListener('click', () => {
+          const column = th.dataset.column;
+          if (sortColumn === column) {
+            sortDirection = sortDirection === 'asc' ? 'desc' : 'asc';
+          } else {
+            sortColumn = column;
+            sortDirection = 'desc';
+          }
+          renderTable();
+        });
+      });
+    }
+
+    // Initialize
+    renderPresetButtons();
+    setupSorting();
+    renderTable();
+
+    // Populate family filter dropdown
+    const families = [...new Set(Object.values(MODELS).map(m => m.family))].sort();
+    const familySelect = document.getElementById('family-filter');
+    families.forEach(f => {
+      const opt = document.createElement('option');
+      opt.value = f;
+      opt.textContent = f;
+      familySelect.appendChild(opt);
+    });
+
+    // Family filter handler
+    familySelect.addEventListener('change', (e) => {
+      currentFamily = e.target.value;
+      renderTable();
+    });
+
+    // TP/PP input handlers
+    document.getElementById('tp-input').addEventListener('input', (e) => {
+      const value = parseInt(e.target.value);
+      if (value >= 1 && value <= 16) {
+        currentTpSize = value;
+        renderTable();
+      }
+    });
+
+    document.getElementById('pp-input').addEventListener('input', (e) => {
+      const value = parseInt(e.target.value);
+      if (value >= 1 && value <= 16) {
+        currentPpSize = value;
+        renderTable();
+      }
+    });
+  </script>
+</body>
+</html>