license: other
license_name: lfm-open-license-v1.0
license_link: https://huggingface.co/LiquidAI/LFM3-Omni-2B/blob/main/LICENSE
base_model:

• LiquidAI/LFM3-Omni-2B
• mizahnyx/lfm3-omni-2b-abliterated-heretic-v2
  library_name: gguf
  pipeline_tag: any-to-any
  tags:
• gguf
• liquid
• lfm3
• lfm3-omni
• omni
• multimodal
• heretic
• abliterated
• uncensored
• decensored
• llama.cpp
  language:
• en
• ar
• zh
• fr
• de
• it
• ja
• ko
• pt
• es
• vi
• th
• id
• hi
• ru
• pl
• da
• el
• fi
• he
• nl
• sv
• tr

LFM3-Omni-2B abliterated — GGUF (Q8_0 / Q4_K_M / Q2_K)

GGUF builds of mizahnyx/lfm3-omni-2b-abliterated-heretic-v2,
a refusal-ablated derivative of LiquidAI/LFM3-Omni-2B
made with Heretic v2.0.0 using the ARA
modifier (60 Optuna trials).

For the full-precision safetensors weights, config, and tokenizer, use the
safetensors repo.

⚠️ MODIFICATION NOTICE

These files are modified derivatives. The text-backbone weights were
permanently altered by automated refusal ablation. The vision encoder, audio
encoder, and both projectors were not modified — every mmproj-* file here
is interchangeable with the upstream base model's.

Distributed under the same LFM Open License v1.0 as the base model
(LICENSE). Not affiliated with or endorsed by Liquid AI.

Files

The K-quants (Q4_K_M, Q2_K) were produced with the included importance matrix,
generated from this model on wikitext-2. It is published so you can reproduce the
quants or make your own.

Quantization quality

The projector is untouched by the ablation; only the text model is.

All three preserve the ablation. But Q2_K at 3.43 BPW is genuinely lossy — it
passes these easy probes and will degrade on harder reasoning. Treat it as a
fit-on-small-hardware tier, not a general replacement.

Usage

Requires a recent llama.cpp build; LFM3 support is new.

# text only
llama-cli -m LFM3-Omni-2B-abliterated-Q4_K_M.gguf -p "Your prompt" -n 256

# with image / audio / video input
llama-server -m LFM3-Omni-2B-abliterated-Q4_K_M.gguf \
             --mmproj mmproj-LFM3-Omni-2B-F16.gguf --jinja

If a client reports an unknown architecture, update llama.cpp — the LFM3
converter and runtime (conversion/lfm3.py, src/models/lfm3.cpp) are recent
additions.

Re-quantizing

llama-quantize --imatrix imatrix/LFM3-Omni-2B-abliterated.imatrix.gguf \
  <a higher-precision gguf> <output.gguf> Q5_K_M

Model details

Heretic's metrics for the released trial: 1/100 keyword refusals, 0/100
classifier refusals, KL 0.063 vs the base model's 99/100 and 98/100.

A trial with 0/100 refusals was rejected: at KL 2.27 it was incoherent (it
answered 17 * 23 with "Ineq 17 23"). See the full report in the safetensors
repo.

Intended use and limitations

Research and red-teaming only. No safety filtering; it will comply with
harmful requests. Not suitable for production or user-facing deployment.

Not evaluated on standard benchmarks (MMLU, GSM8K, etc.) — only a 12-probe
spot-check plus Heretic's internal metrics. Multimodal output quality after
ablation was not measured. Inherits all base-model limitations.

License

LFM Open License v1.0 — a copy is included as LICENSE. Redistribution
permitted with attribution; commercial use is restricted (§5). No trademark
rights granted; "LFM"/"Liquid AI" are trademarks of Liquid AI, Inc., used here
only to describe the origin of the work.