jfan/gemma-4-e2b-web-mtp-vision-litert-lm

Gemma 4 E2B — web-optimized, MTP + vision (${SIZE_GIB:-~2.3} GB)

Web-optimized gemma-4-E2B-it LiteRT-LM bundle for browser (WebGPU) use, repackaged with litert-lm-builder from litert-community/gemma-4-E2B-it-litert-lm:

• Web-optimized core: tf_lite_artisan_text_decoder (backend constraint gpu_artisan) — the same WebGPU-compiled decoder as gemma-4-E2B-it-web.litertlm, with per-layer embedding weights inlined (no separate embedder sections, lower peak memory in the browser).
• MTP: tf_lite_mtp_drafter — Multi-Token Prediction speculative decoding (enable via enableSpeculativeDecoding / enable_speculative_decoding).
• Vision: tf_lite_vision_encoder (fp16) + tf_lite_vision_adapter
  • tf_lite_end_of_vision — image input support (the official web bundle is text-only; this one is not).
• Chat template (jinja) + Gemma-4 LLM metadata embedded in the container.
• Audio sections dropped (the web runtime has no audio path).

Run

bash pip install -U litert-lm litert-lm run --from-huggingface-repo jfan/gemma-4-e2b-web-mtp-vision-litert-lm gemma-4-E2B-it-web-mtp-vision.litertlm --prompt="Hello"

In the web app this file is the MODEL_URL target (loaded with @litert-lm/core, default Backend.GPU_ARTISAN).