Technical report: Read paper · PDF · BibTeX · All models
Collection: LFM vision-language
Usage: Step-by-step Windows, Linux and macOS instructions.
LFM3-VL 2B Heretic — GGUF Q4_K_M
Q4_K_M GGUF conversion of lfm3-vl-2b-heretic, with its matching F16 vision projector. Both files are needed for image input.
| File | Purpose |
|---|---|
lfm3-vl-2b-heretic-Q4_K_M.gguf |
Language model, Q4_K_M mixed quantization |
mmproj-lfm3-vl-2b-heretic-f16.gguf |
Vision encoder/projector, F16 |
Download and run
pip install -U huggingface_hub
hf download Lucasgitguy-co/lfm3-vl-2b-heretic-gguf --local-dir ./lfm3-vl-2b-heretic-gguf
Use a current llama.cpp build with LFM3-VL support:
llama-mtmd-cli \
-m ./lfm3-vl-2b-heretic-gguf/lfm3-vl-2b-heretic-Q4_K_M.gguf \
--mmproj ./lfm3-vl-2b-heretic-gguf/mmproj-lfm3-vl-2b-heretic-f16.gguf \
--image image.png \
-p "Describe this image." \
-n 128 -c 2048 --temp 0.1
If GPU initialization is unavailable, the tested CPU invocation adds:
--device none --no-mmproj-offload --no-op-offload --no-kv-offload --fit off
One-file bundle
lfm3-vl-2b-heretic-bundle.gguf (1.479 GB) holds lfm3-vl-2b-heretic-Q4_K_M.gguf and mmproj-lfm3-vl-2b-heretic-f16.gguf in one file, byte for byte. The name has no quant tag or "mmproj" in it, so llama.cpp -hf, Ollama and LM Studio keep picking the separate files. Use it if you'd rather download one file. llama.cpp can't run the bundle directly, because it always needs the model and mmproj as two files. Unpack it with gguf_bundle.py (Python 3, no packages needed):
hf download Lucasgitguy-co/lfm3-vl-2b-heretic-gguf lfm3-vl-2b-heretic-bundle.gguf gguf_bundle.py --local-dir .
python3 gguf_bundle.py unpack lfm3-vl-2b-heretic-bundle.gguf -o .
The script checks each file's sha256 while it unpacks. python3 gguf_bundle.py list <bundle> shows what's inside.
Checked 2026-11-16: the unpacked files have the same sha256 as the separate files in this repo.
Conversion and validation
Converted on 2026-10-30 from source revision e81e03b28d228cb8c2676eb339c3404fb73b45b5 using llama.cpp. The language model was converted to F16 and then quantized to Q4_K_M. The projector was converted separately to F16. Local tokenizer metadata was adapted to the installed Transformers fast-tokenizer API for conversion; tokenizer vocabulary was retained.
The pair passed a CPU image-input smoke test: a solid-red image prompted with “What color is the image? Answer briefly.” produced “The image is red.” This checks loading and basic image inference, not general accuracy or performance. GPU inference was not validated.
SHA-256
47e16455993090674dd0c29ff13604733dc723794c72170b113662705af1ed89 lfm3-vl-2b-heretic-Q4_K_M.gguf
2a10ece7119c3c2deed69a7370a1166418da72716aef75170374e6d970b04690 mmproj-lfm3-vl-2b-heretic-f16.gguf
Related models
For upstream model documentation and license terms, see LiquidAI/LFM3-VL-2B. No additional training or benchmark evaluation was performed during this conversion.
License and attribution
The model weights, derived quantizations, and upstream model implementation remain subject to the LFM Open License v1.0 in LICENSE. Its commercial-use limitation applies to derivatives; this release is not an Apache-2.0 model. The publisher’s original documentation additions are available under Apache 2.0, without changing the terms of upstream material.