diff --git a/docs/conf.py b/docs/conf.py
index 9993b767..528a3c47 100644
--- a/docs/conf.py
+++ b/docs/conf.py
@@ -54,6 +54,10 @@
         'gemma/peft/README.md': 'peft.md',
     },
     # Redirect the empty `XX.html` pages to their `colab_XX.html`
-    redirects={name: f'colab_{name}.html' for name in _COLABS_NAMES},
+    redirects={
+        name: f'colab_{name}.html'
+        for name in _COLABS_NAMES
+        if name != 'tokenizer'  # tokenizer.md has standalone content
+    },
     globals=globals(),
 )
diff --git a/docs/index.md b/docs/index.md
index ed5a18e8..1060b954 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -10,6 +10,7 @@
     colab_tool_use
     colab_sharding
     colab_tokenizer
+    tokenizer
     colab_lora_sampling
     colab_lora_finetuning
     peft
diff --git a/docs/tokenizer.md b/docs/tokenizer.md
index ea74e693..90797de9 100644
--- a/docs/tokenizer.md
+++ b/docs/tokenizer.md
@@ -1 +1,129 @@
 # Tokenizer
+
+Gemma tokenizers convert text into token IDs that the model can process. Each Gemma version has its own tokenizer with different vocabulary and special tokens.
+
+For a step-by-step tutorial, see the [tokenizer Colab](colab_tokenizer.ipynb).
+
+## Quick start
+
+```python
+from gemma import gm
+
+tokenizer = gm.text.Gemma3Tokenizer()
+
+# Encode text to token IDs
+tokenizer.encode('Derinkuyu is an underground city.')
+# [8636, 979, 78904, 603, 671, 30073, 3413, 235265]
+
+# Split text into token pieces
+tokenizer.split('Derinkuyu is an underground city.')
+# ['Der', 'ink', 'uyu', ' is', ' an', ' underground', ' city', '.']
+
+# Decode token IDs back to text
+tokenizer.decode([8636, 979, 78904, 603, 671, 30073, 3413, 235265])
+# 'Derinkuyu is an underground city.'
+```
+
+## Tokenizer versions
+
+Each Gemma model family has a corresponding tokenizer:
+
+```python
+tokenizer = gm.text.Gemma2Tokenizer()
+tokenizer = gm.text.Gemma3Tokenizer()
+tokenizer = gm.text.Gemma3nTokenizer()
+```
+
+You can also create a tokenizer by version number:
+
+```python
+tokenizer = gm.text.Tokenizer.from_version(3)
+```
+
+Feature               | Gemma 2 | Gemma 3   | Gemma 3n
+--------------------- | ------- | --------- | --------
+Vocab size            | 256,000 | 256,000   | 256,000
+Image tokens          | No      | Yes       | Yes
+Tool tokens           | No      | Yes       | Yes
+
+## Special tokens
+
+Special tokens are accessible via `tokenizer.special_tokens`:
+
+```python
+tokenizer = gm.text.Gemma3Tokenizer()
+tokenizer.special_tokens.BOS   # 2
+tokenizer.special_tokens.EOS   # 1
+```
+
+Token              | Gemma 2 | Gemma 3 / 3n | Description
+------------------ | ------- | ------------ | -----------
+`PAD`              | 0       | 0            | Padding
+`EOS`              | 1       | 1            | End of sentence
+`BOS`              | 2       | 2            | Begin of sentence
+`UNK`              | 3       | 3            | Unknown
+`MASK`             | 4       | 4            | Mask
+`CUSTOM`           | 7       | 6            | Start of custom token range
+`START_OF_TURN`    | 106     | 105          | `<start_of_turn>`
+`END_OF_TURN`      | 107     | 106          | `<end_of_turn>`
+`START_OF_IMAGE`   | —       | 255999       | `<start_of_image>` (Gemma 3+ only)
+`END_OF_IMAGE`     | —       | 256000       | `<end_of_image>` (Gemma 3+ only)
+
+### BOS / EOS
+
+The `<bos>` token should appear once at the beginning of the input. You can add it with `add_bos=True`:
+
+```python
+tokenizer.encode('Hello world!', add_bos=True)
+# [2, 4521, 2134, 235341]
+```
+
+Similarly, `add_eos=True` appends the end-of-sentence token.
+
+### Turn tokens
+
+Instruction-tuned models use `<start_of_turn>` / `<end_of_turn>` to separate user and model turns:
+
+```python
+tokenizer.encode("""<start_of_turn>user
+Knock knock.<end_of_turn>
+<start_of_turn>model
+Who's there?<end_of_turn>""")
+```
+
+### Image token
+
+In Gemma 3, `<start_of_image>` marks where an image should be inserted in the prompt. The model expands this internally into soft image tokens.
+
+## Custom tokens
+
+All Gemma tokenizers reserve 99 unused token slots (IDs `CUSTOM + 0` through `CUSTOM + 98`) that can be mapped to custom strings:
+
+```python
+tokenizer = gm.text.Gemma3Tokenizer(
+    custom_tokens={
+        0: '<my_custom_tag>',
+        17: '<my_other_tag>',
+    },
+)
+
+tokenizer.encode('<my_other_tag>')  # [24]
+tokenizer.decode(tokenizer.special_tokens.CUSTOM + 17)  # '<my_other_tag>'
+```
+
+## Important notes
+
+**Whitespace is part of tokens.** The model treats ` hello` (with leading space) and `hello` as different tokens with different IDs:
+
+```python
+tokenizer.encode(' hello')  # [25612]
+tokenizer.encode('hello')   # [17534]
+```
+
+**Avoid trailing spaces.** When preparing prompts for next-token prediction, a trailing space creates an unusual final token that may degrade quality:
+
+```python
+tokenizer.split('The capital of France is ')
+# ['The', ' capital', ' of', ' France', ' is', ' ']
+#                                               ^^^ unusual trailing token
+```