Run clef-flash, Cloudflare's open-weight decision model, locally on a Mac and turn your webcam into typed decisions with probabilities.

A decision model doesn't generate text. You give it a state (here, a webcam frame) and a schema of questions, and it returns a probability for every allowed answer in one forward pass. Read the announcement.

You need an Apple Silicon Mac with 32 GB+ of memory (the model is ~19 GB) and uv.

uv sync
uv run hf download Cloudflare/clef-flash --local-dir models/clef-flash
uv run uvicorn server:app --port 8787

Open http://localhost:8787 and allow camera access.

Questions use the same schema as @cf/cloudflare/clef on Workers AI (and Jev / SystemOne):

{
  "wearing_glasses": { "type": "noul", "instructions": "Wearing glasses?" },
  "mood": { "type": "choice", "criteria": { "happy": null, "neutral": null, "tired": null } },
  "energy": { "type": "score", "criteria": ["low", "medium", "high"] }
}

• noul: yes or no
• choice: pick one option. Descriptions are optional (null)
• score: an ordered scale

Latency grows with prompt length, so short instructions and few options keep it snappy.

Measured on an M5 Max with uv run python bench.py, which also checks that the optimized model gives the same answers as the reference:

On a Mac, Qwen3.5's linear-attention layers fall back to a slow reference implementation, because the fast kernels are CUDA-only. Most of that time goes to torch.linalg.solve_triangular. mps_kernels.py computes the same thing with batched block inversion, which Apple GPUs handle well. It is patched in automatically on MPS and CPU, and left alone on CUDA.

joint_schema_model.py (Clef's reference inference code) is downloaded with the model.

This code is Apache 2.0 licensed, and so is the clef-flash model.