Lumix-4B is a compact, vision-capable coding and reasoning model from Blackfrost, distributed as GGUF quantizations for llama.cpp and Ollama. It is a fine-tune of Qwen3.5-4B, tuned to drive the Shadow coding agent well above its size.

Base: Qwen/Qwen3.5-4B (GGUF architecture qwen35)
Parameters: ~4B
Modalities: text + vision (image input via the bundled mmproj)
Context: long-context, agentic coding — one clean tool call at a time, look → act → verify

Files

File	Quant	Notes
`Lumix-4B-Q3_K_M.gguf`	Q3_K_M	smallest
`Lumix-4B-Q4_K_M.gguf`	Q4_K_M	recommended
`Lumix-4B-Q5_K_M.gguf`	Q5_K_M
`Lumix-4B-Q6_K.gguf`	Q6_K
`Lumix-4B-Q8_0.gguf`	Q8_0	near-lossless
`Lumix-4B-f16.gguf`	F16	full precision
`Lumix-4B-mmproj.gguf`	—	vision projector (pair with any quant for image input)

Usage

llama.cpp

bash

llama-cli -hf Blackfrost-AI/Lumix-4B:Q4_K_M
# vision: also pass --mmproj Lumix-4B-mmproj.gguf and an --image

Ollama

ollama run hf.co/Blackfrost-AI/Lumix-4B:Q4_K_M

License

Released under the Apache-2.0 license.