LFM3-VLD-2B GGUF
GGUF exports of webbrain-one/lfm3-vld-2b for running in llama.cpp both as a multimodal chat model (/v1/chat/completions) and as a multipass decision model (/v1/systemone).
Features untied embeddings (token_embd.weight and output.weight exported separately, 288 tensors), restoring original LFM3-VL-2B generative language head on top of the d1-2B decision backbone.
Files
| File | Quantization | Size | Description |
|---|---|---|---|
LFM3-VL-2B-VLDR-BF16.gguf |
BF16 |
3.8 GB | Full BF16 precision with untied head |
LFM3-VL-2B-VLDR-Q8_0.gguf |
Q8_0 |
2.0 GB | 8-bit quantized weights, high fidelity |
LFM3-VL-2B-VLDR-Q4_K_M.gguf |
Q4_K_M |
1.2 GB | 4-bit medium k-quant, light and fast |
mmproj-LFM3-VL-2B-VLDR-F16.gguf |
Vision Projector (F16) |
362 MB | Multimodal projector for vision inputs |
Usage with llama-server
llama-server \
-m LFM3-VL-2B-VLDR-Q8_0.gguf \
--mmproj mmproj-LFM3-VL-2B-VLDR-F16.gguf \
-ngl 99 -c 16384 --port 8080