library_name: gguf
pipeline_tag: any-to-any
license: other
license_name: lfm1.0
license_link: LICENSE
base_model: NeosQ/LFM3-Omni-2B-Uncensored-abliterated
tags:
- gguf
- llama.cpp
- lfm3
- lfm3-omni
- abliteration
- uncensored
- refusal-removal
- audio
- vision
- speech
- q4_k_m
- q8_0
LFM3-Omni-2B-Uncensored-abliterated.gguf
Research-only, no refusal guardrail. This is an abliterated checkpoint: it will answer
requests the base model declines and can produce unsafe or illegal content. Not for
production or consumer-facing use. You are responsible for what you generate.
GGUF conversion of NeosQ/LFM3-Omni-2B-Uncensored-abliterated, which isLiquidAI/LFM3-Omni-2B with its refusal direction removed using Abliteration (MPOA, direction ablation).
Three quantisation levels plus the multimodal projector, all in one repo.
Text backbone and projector are separate files. The *gguf main file is text-only on its
own; pass the mmproj-* file with -mm / --mmproj to get image input, audio input and speech
output. Both are required for anything multimodal.
Files
| File | Size | Tensors | Notes |
|---|---|---|---|
LFM3-Omni-2B-Uncensored-abliterated-BF16.gguf |
3.2 GiB | 287 | unquantised export, reference file |
LFM3-Omni-2B-Uncensored-abliterated-Q8_0.gguf |
1.9 GiB | 287 | 8-bit, token embedding kept in F16 |
LFM3-Omni-2B-Uncensored-abliterated-Q4_K_M.gguf |
1.0 GiB | 287 | stock k-quant, the one to grab for laptops |
mmproj-LFM3-Omni-2B-Uncensored-abliterated-F16.gguf |
905.8 MiB | 458 | vision + audio encoder, depthformer and codec — required for image/audio in and speech out |
Q4_K_M is the default pick: ~1 GB and it keeps the text quality you would expect at 4-bit.
Take Q8_0 if you have the RAM and want the closest match to the BF16 export. BF16 is
included as the exact reference the others were quantised from.
Requirements
A llama.cpp build with LFM3 / LFM3-Omni support. These files were produced and tested at
build b1-09338b8; older builds fail with unknown model architecture: 'lfm3'.
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
Usage
Text only:
llama-cli -m LFM3-Omni-2B-Uncensored-abliterated-Q4_K_M.gguf -p "Explain what a mixture-of-experts layer does." -n 256
Image, audio, video or speech output (projector required):
llama-mtmd-cli -m LFM3-Omni-2B-Uncensored-abliterated-Q4_K_M.gguf \
--mmproj mmproj-LFM3-Omni-2B-Uncensored-abliterated-F16.gguf \
--image photo.jpg -p "Describe the image in one sentence."
llama-mtmd-cli -m LFM3-Omni-2B-Uncensored-abliterated-Q4_K_M.gguf --mmproj mmproj-LFM3-Omni-2B-Uncensored-abliterated-F16.gguf \
--audio speech.wav -p "" --temp 0 -sys "Perform ASR in English."
Server:
llama-server -m LFM3-Omni-2B-Uncensored-abliterated-Q4_K_M.gguf --mmproj mmproj-LFM3-Omni-2B-Uncensored-abliterated-F16.gguf --port 8080
The model is non-reasoning; --reasoning off keeps llama-server from wrapping answers.
Long-form output
This is a 2B model and long answers are its weak spot. With llama.cpp defaults it loops,
drifts into code or collapses into an outline. The settings that fixed it on the transformers
build (repetition_penalty=1.3 over a 1024-token window, temperature=1.0, top_p=0.95) map
to:
llama-cli -m LFM3-Omni-2B-Uncensored-abliterated-Q4_K_M.gguf --temp 1.0 --top-p 0.95 \
--repeat-penalty 1.3 --repeat-last-n 1024 -n 1600
--repeat-last-n is the equivalent of the model's text_repetition_window; the default 64 is
shorter than one sentence, which is why paragraphs get copied verbatim. One pass reliably
holds 800-2,000 characters; longer text needs section-by-section generation. The full
measurements are in the transformers model card.
How these were made
| Step | Tool | Result |
|---|---|---|
| Convert backbone | convert_hf_to_gguf.py --outtype bf16 |
BF16, 287 tensors |
| Convert projector | convert_hf_to_gguf.py --outtype f16 --mmproj |
F16, 458 tensors |
| Q8_0 | llama-quantize --token-embedding-type F16 |
8-bit, embedding untouched |
| Q4_K_M | llama-quantize stock preset |
no imatrix, no custom mix |
No retraining and no calibration data: these are direct quantisations of the BF16 export.
Multimodal was verified end-to-end (llama-mtmd-cli with an image prompt) and all four files
load in llama-cli; per-file SHA-256:
0200dd9e6b5fddcd5c97a0727d5533cbb268fc11b1c31ec1951acf07194476d7 LFM3-Omni-2B-Uncensored-abliterated-BF16.gguf
90d515472a725563c53f6f5644b0a14747010c6a18a2362787917b421f67daca LFM3-Omni-2B-Uncensored-abliterated-Q8_0.gguf
8dc18c47bac848a84070b074ce232a1923d3c2bfa2252fb6ee4c0d72a112baf9 LFM3-Omni-2B-Uncensored-abliterated-Q4_K_M.gguf
7f8ef74353b459662cdbe20f180a19a402e81a4e175121044cd9ddd555072d02 mmproj-LFM3-Omni-2B-Uncensored-abliterated-F16.gguf
License
LFM Open License v1.0, inherited from LiquidAI/LFM3-Omni-2B (see LICENSE). All of its
terms, including the commercial-use threshold, apply to this derivative and to these GGUF
files. Community conversion, not affiliated with or endorsed by Liquid AI.