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
library_name: transformers
pipeline_tag: any-to-any
language:

• en
• ar
• zh
• fr
• de
• it
• ja
• ko
• pt
• es
• vi
• th
• id
• hi
• ru
• pl
• da
• el
• fi
• he
• nl
• sv
• tr
  tags:
• liquid
• lfm3
• lfm3-omni
• omni
• multimodal
• heretic
• abliterated
• uncensored
• decensored
• gguf

LFM3-Omni-2B — abliterated (Heretic v2, ARA)

A refusal-ablated derivative of LiquidAI/LFM3-Omni-2B,
produced with Heretic v2.0.0 using the ARA
(Arbitrary-Rank Ablation) modifier.

⚠️ MODIFICATION NOTICE

This model has been modified. The text-backbone weights of the base model
were permanently altered by an automated refusal-ablation procedure (Heretic v2,
ARA, 60 Optuna trials, seed 42). The weights in model.safetensors differ
from the base checkpoint (base sha256 76c0ec2c…d2763; this model sha256
375b950f…846a).

The vision encoder, audio encoder, and both projectors were not modified —
only attn.o_proj and mlp.down_proj in the 24 text-backbone layers.

This is a derivative work distributed under the same LFM Open License v1.0 as
the base model. It is not affiliated with, endorsed by, or supported by
Liquid AI. See LICENSE.

What this is

LFM3-Omni-2B is Liquid AI's omni-modal model (text, image, video and audio in;
text and speech out). This release removes its learned refusal behaviour so the
model answers requests the base model declines.

Zero refusals is not the goal by itself — a coherent model is. Heretic ranks
trials by fewest refusals first, and that ranking point (0/100 refusals, KL 2.27)
was measurably degraded: it answered 17 * 23 with "Ineq 17 23". The released
weights are Pareto index 2, chosen because it keeps refusal removal and the
base model's capability.

Method

Evaluation

Heretic's own metrics, on the study's held-in prompt sets:

Independent spot-check, greedy decoding, same prompts for both models:

Capability therefore matches the base model on all five probes while refusing
none of the seven harmful ones.

Quantized variants

GGUF builds are also published separately at mizahnyx/lfm3-omni-2b-abliterated-heretic-v2-GGUF, which is the repository to use with llama.cpp / LM Studio. The copies under
gguf/ here are the same files.

Produced with llama.cpp; the K-quants use an importance matrix generated from
this model (wikitext-2, 100 chunks) — no usable imatrix existed upstream.

All four text quants answer 391, Canberra and the harmful probe correctly.
Q4_K_M is the recommended trade-off. Q2_K at 3.43 BPW is a last-resort tier —
it passes these easy probes but will visibly degrade on harder tasks.

The projector is unchanged from the base model, so mmproj-* files are
interchangeable with the upstream ones.

Usage

Transformers

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "mizahnyx/lfm3-omni-2b-abliterated-heretic-v2"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
).eval()

conversation = [
    {"role": "system", "content": "Respond with text."},
    {"role": "user", "content": "Explain how acoustic resonance works in musical instruments."},
]
inputs = processor.apply_chat_template(
    conversation, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, generation_mode="text")

text = processor.decode(
    outputs.sequences[0, inputs["input_ids"].shape[1]:outputs.sequence_lengths[0]],
    skip_special_tokens=True,
)
print(text)

Note the non-standard generation_mode="text" argument and the
sequence_lengths-based decoding — both are LFM3-Omni specifics.

llama.cpp

Requires a recent llama.cpp build (LFM3 support is new).

llama-cli -m LFM3-Omni-2B-abliterated-Q4_K_M.gguf \
          --mmproj mmproj-LFM3-Omni-2B-F16.gguf \
          -p "Your prompt"

For image/video/audio input, pass the mmproj file as shown; for text-only use
you can omit it.

Intended use and limitations

Intended use: safety research, red-teaming, evaluation of refusal-ablation
techniques, and research into model alignment failure modes.

This model has no safety filtering and will comply with harmful requests. It
is unsuitable for any user-facing or production deployment, and unsuitable for
use by or around minors. You are responsible for how you use it.

Known limitations:

• Abliteration is a blunt instrument. It removes refusals across the board,
  including refusals that were appropriate. Expect it to be worse than the base
  model at recognising genuinely harmful edge cases, and to be less cautious on
  borderline requests that the base model handled well.
• Capability was verified only on a small 12-probe spot-check plus Heretic's
  internal metrics. It has not been evaluated on standard benchmarks (MMLU,
  GSM8K, etc.), and no multimodal output quality was measured after ablation.
• Q2_K specifically is lossy enough to change behaviour on non-trivial tasks.
• Inherits all base-model limitations, including modality-specific failure modes,
  language coverage gaps, and hallucination.

Reproducibility

Seeds, prompt sets, and every setting are recorded above; the full run report
(with the five patches required to run Heretic on this architecture, the complete
Pareto front, and the evaluation transcripts) is in docs/REPORT.md.

The ablation is reproducible from the unmodified base checkpoint with Heretic
v2.0.0, seed = 42, n_trials = 60, n_startup_trials = 20, ARA, and the three
scorers listed above.

License

Distributed under the LFM Open License v1.0, the same license as the base
model. A copy is included as LICENSE. Key points: redistribution is
permitted with attribution; commercial use is restricted (see §5 —
entities above the revenue threshold are not licensed); trademark rights are not
granted.

This derivative is not affiliated with or endorsed by Liquid AI. "LFM" and
"Liquid AI" are trademarks of Liquid AI, Inc., used here only to describe the
origin of the work.