Gemma 4 E2B for A2UI Express, WebLLM package (retrain 2026-11-07)
The same adapter as soverius-ai/gemma-4-E2B-it-A2UI-Express-ONNX (checkpoint
540, SHA-256 54c7414491e36610133c5f22e0397a66e97d901e3b119238abf50042dc7aa323)
merged into the BF16 base and converted to MLC q4f16_1 by patching the
official mlc-ai/gemma-4-E2B-it-q4f16_1-MLC conversion (240 groups, 445
arrays), packaged with the WebLLM 0.2.87 runtime (model.wasm) and the
a2ui-mlc-manifest.json descriptor the demo verifies. Model id
gemma-4-e2b-express-54c7414491e3-q4f16_1-MLC.
Note: in the demo's tests this q4f16_1 path was both slower (about 35 s to
the first token for the 6.3k-token prompt on an M5 Max) and less faithful than
the ONNX export; the MLX 4-bit and BF16 forms of the same weights compile every
test prompt while this package flips tokens on some turns. Prefer the ONNX
repository. Licenses: Apache-2.0 for the MLC base conversion and TVM runtime
(LICENSE, NOTICE.txt), Gemma terms for the weights.