README.md

metadata

title: BlackRiver Gemma 4 Multimodal
emoji: 🌊
colorFrom: blue
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
short_description: Gemma 4 E2B/E4B with vision, audio and text running locally on WebGPU.
models:
  - onnx-community/gemma-4-E2B-it-ONNX
  - onnx-community/gemma-4-E4B-it-ONNX

BlackRiver AI · Gemma 4 Multimodal

A standalone BlackRiver AI implementation of Gemma 4 for fully local browser inference with WebGPU.

Capabilities

• Gemma 4 E2B and E4B model selection
• Text conversation
• Image understanding
• Audio recording and transcription/instruction input
• Streaming generation
• Local browser cache for model weights
• Rendered WebGPU kernel inspection
• No inference server: prompts and media stay on the user's device

The first model load downloads the selected ONNX weights from Hugging Face. A WebGPU-capable browser and sufficient GPU memory are required. E4B requires more memory than E2B.

BlackRiver AI implementation

This Space packages the models into a unified BlackRiver interface and adds the multimodal browser workflow used here, including image attachment, microphone capture, 16 kHz audio preparation, E2B/E4B selection and media-history handling for reliable repeated generations.

BlackRiver AI does not claim authorship of the underlying models or the original WebGPU research work.

Credits

• Gemma 4 models: Google — Gemma 4 E2B
• ONNX model conversions: ONNX Community E2B and E4B
• Browser runtime: Transformers.js
• Original WebGPU demo foundation: WebML Community
• Original optimized WebGPU kernels: Fable 5, as credited by the upstream demo

Model use remains subject to the licenses and terms published in the respective model repositories.

Files

• index.html — complete BlackRiver multimodal interface
• transformers-gemma4.js — Gemma 4 E2B/E4B Transformers.js adapter
• landing.js — landing-page WebGL background

No build step or server backend is required.