Add Gemma 3n support (E2B, macOS)

models/gemma3n/README.md

@@ -0,0 +1,38 @@
+# Gemma 3n
+
+Google's Gemma 3n for on-device inference via Core AI.
+
+## Supported Models
+
+| Model           | Parameters          | Context | macOS | iOS |
+| --------------- | ------------------- | ------- | ----- | --- |
+| gemma-3n-E2B-it | ~5B (E2B effective) | 32768   | Yes   | No  |
+
+## Export
+
+```bash
+uv run coreai.llm.export google/gemma-3n-E2B-it
+```
+
+## Run
+
+```bash
+swift run -c release llm-runner --model path/to/exported_model --prompt "Hello"
+```
+
+## Benchmark
+
+```bash
+swift run -c release llm-benchmark --model path/to/exported_model -p 512 -g 1024 -n 5
+```
+
+## Evaluation
+
+Perplexity on WikiText-2 and accuracy on tinyMMLU, computed with the Core AI PyTorch models (4K context).
+
+| Compression      | Wikitext PPL | tinyMMLU acc |
+| ---------------- | ------------ | ------------ |
+| none (float16)   | 61.53        | 58.75%       |
+| 4-bit quantized  | 72.13        | 59.27%       |
+
+High wikitext PPL is expected for instruction-tuned models — the probability distribution is optimized for dialogue, not raw text. tinyMMLU accuracy is the appropriate quality metric.

python/src/coreai_models/export/metadata.py

@@ -95,6 +95,15 @@ class AIModelMetadataFields:
             "Source: https://huggingface.co/google/gemma-3-12b-it"
         ),
     ),
+    "google/gemma-3n-E2B-it": AIModelMetadataFields(
+        author="Gemma Team",
+        license="Gemma Terms of Use",
+        model_description=(
+            "Gemma 3n E2B is a ~5B-parameter on-device language model from Google's "
+            "Gemma 3n family using AltUp for efficient inference. "
+            "Source: https://huggingface.co/google/gemma-3n-E2B-it"
+        ),
+    ),
     "mistralai/Mistral-7B-Instruct-v0.3": AIModelMetadataFields(
         author="Mistral AI",
         license="Apache-2.0",

python/src/coreai_models/model_registry.py

@@ -174,6 +174,16 @@ class UtilityModel:
         "float16",
         131072,
     ),
+    ModelPreset(
+        "gemma-3n-e2b-it",
+        "google/gemma-3n-E2B-it",
+        "gemma3n",
+        "llm",
+        "macOS",
+        "4bit",
+        "float16",
+        32768,
+    ),
     # --- iOS (compression = palettized) ---
     ModelPreset(
         "qwen3-0.6b",