LFM3-VL-2B-MLX-8bit-uzu
LiquidAI/LFM3-VL-2B-MLX-8bit repacked for the Uzu / Mirai inference engine on Apple Silicon.
Conversion
Text decoder: repack only — MLX affine 8-bit (group 64) tensors carried over as-is (code words U32→U8, scale/bias planes transposed for OutputInput linears, norms BF16→F32, fused qkvg_projection [Q|K|V|GATE] and gate|up MLP layouts).
Vision tower: packed BF16 as-is into vision.safetensors, run by uzu's Metal vision pipeline; vision_config.json carries patchify/merge and projector parameters.
Files
config.json— uzu language model and decoder configuration.vision_config.json— vision tower architecture, patch embedding, projector.encoding.json— chat encoding pointer.tokenizer.json— tokenizer.model.safetensors— quantized text weights (~1.9 GB).vision.safetensors— vision tower weights (~379 MB).
Usage
cli server --model LFM3-VL-2B-MLX-8bit-uzu --port 8000
OpenAI-compatible multimodal requests (/v1/chat/completions with image image_url parts, streaming SSE supported).
curl http://127.0.0.1:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "LFM3-VL-2B-MLX-8bit-uzu",
"messages": [{"role": "user", "content": [
{"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}},
{"type": "text", "text": "Describe this image."}
]}]
}'