welcoma/gemma-4-E4B-it-q4f16_1-MLC

Gemma 4 E4B IT q4f16_1 MLC

Custom text-first MLC/WebLLM packaging of google/gemma-4-E4B-it in q4f16_1 for browser-local WebGPU inference.

Status

This is a build candidate, not an official mlc-ai release. The compile path has completed successfully; browser runtime validation still needs a WebGPU device that exposes shader-f16.

Build Summary

• Base model: google/gemma-4-E4B-it
• Quantization: q4f16_1
• Runtime target: webgpu
• Model type: gemma4
• Conversation template: gemma4_instruction
• Context window: 4096
• Prefill chunk size: 512
• Sliding window: 512
• Quantized parameters: 7,996,157,418
• Parameter size after quantization: 3.976 GB
• Build VM: gemma4-qwik-e4b-builder (m3-megamem-128, europe-west1-b)
• MLC-LLM commit: 22fe4b7e2e68ff00c12c2069de2060bce3cfe62d
• TVM commit: e96bc0525fb6d59229d40c5a6eb03cde04bb5ed4

WebLLM Usage

import { CreateMLCEngine } from "@mlc-ai/web-llm";

const repo = "https://huggingface.co/welcoma/gemma-4-E4B-it-q4f16_1-MLC";
const appConfig = {
  model_list: [{
    model: `${repo}/resolve/main/`,
    model_id: "gemma-4-E4B-it-q4f16_1-MLC",
    model_lib: `${repo}/resolve/main/libs/gemma-4-E4B-it-q4f16_1-MLC-webgpu.wasm`,
    required_features: ["shader-f16"],
  }],
};

const engine = await CreateMLCEngine("gemma-4-E4B-it-q4f16_1-MLC", { appConfig });

Files

• libs/gemma-4-E4B-it-q4f16_1-MLC-webgpu.wasm: WebGPU model library
• mlc-chat-config.json: MLC runtime configuration
• params_shard_*.bin: quantized parameter shards
• tensor-cache.json: tensor metadata cache
• tokenizer.json, tokenizer_config.json: tokenizer assets
• release-manifest.json: SHA-256 file inventory
• build-provenance.json: build environment and source commit provenance

Limitations

• Text-only packaging. The base model has multimodal components, but image/audio paths are not packaged or validated here.
• Requires browser WebGPU with shader-f16.
• Runtime validation should be performed on a compatible local browser/GPU before treating this as stable.