README: add Footprint & Performance section (34MB app, 1.27GB models, 533 tok/s)

Author: seyeong-han

execuwhisper/macos/README.md

@@ -38,6 +38,26 @@ https://github.com/user-attachments/assets/b840bf99-e221-4c19-ba2e-771903fa357b
 
 (Recording outline lives in [`docs/DEMO_SCRIPT.md`](docs/DEMO_SCRIPT.md).)
 
+## Footprint & Performance
+
+ExecuWhisper is one of the smaller fully-on-device dictation stacks shippable today. Measurements taken on an Apple Silicon Mac during active dictation:
+
+| What | Size |
+|---|---:|
+| App bundle (`ExecuWhisper.app`) — Swift UI + 2 helper binaries + `mlx.metallib` + `libomp.dylib` | **34 MB** |
+| Models on disk (downloaded once on first launch) | **1.27 GB** |
+|    Parakeet-TDT ASR (`model.pte` + tokenizer) | 800 MB |
+|    LFM2.5-350M formatter (`lfm2_5_350m_mlx_4w.pte` + tokenizer) | 472 MB |
+| Peak working-set memory during inference (sum of UI + both helpers) | **~4.8 GB** |
+|    ExecuWhisper UI (SwiftUI) | 106 MB |
+|    `parakeet_helper` (ASR, Metal) | 1.4 GB |
+|    `lfm25_formatter_helper` (LFM2.5, MLX) | 3.3 GB |
+| LFM2.5 formatter throughput (mean over 100-row AMI eval) | **533 tok/s** |
+
+> **For context:** the 34 MB app bundle is roughly 5–10× smaller than a typical Electron-based dictation app, and the 1.27 GB on-disk model footprint is well under what a single 7B-class chat LLM would occupy. Peak memory of ~4.8 GB is the cost of keeping both helper processes warm with their KV caches resident on the Metal GPU; idle steady-state RSS sits closer to ~1.7 GB.
+
+Throughput numbers from `eval/eval_ami_mlx_4w_g32.json` on the [formatter HF repo](https://huggingface.co/younghan-meta/LFM2.5-350M-ExecuWhisper-Formatter); footprint numbers from `vmmap --summary` ("Physical footprint (peak)") on the running app.
+
 ## Architecture
 
 <p align="center">