Decision 3.0 — Core ML (and MLX)
Decision 3.0 by the vLLM Semantic Router team, converted to Core ML for Apple silicon: text, image and video requests on device. Same contract as upstream: one forward pass per question, the 255-way FP32 readout at the last prompt token, a probability for every option, no text generated.
| Folder | Upstream | Size |
|---|---|---|
lite/ |
vllm-sr/d3-lite | 0.8B |
nano/ |
vllm-sr/d3-nano | 2B |
mini/ |
vllm-sr/d3-mini | 4B |
edge/ |
vllm-sr/d3-edge | 0.6B · also ships mlx/ weights and an MLX host (video looks ~4× faster than Core ML) |
Each folder is a full release (text graphs, vision tower, host tables, Python host) with its own README covering files, parity against upstream and speed. Swift: D3VisionManager in FluidUse loads any of them.
License: Apache-2.0, as upstream.
- Downloads last month
- 437
Hardware compatibility
Log In to add your hardware
Quantized
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download FluidInference/d3-coreml --local-dir d3-coreml