LFM3-VL-2B-heretic

A decensored variant of LiquidAI/LFM3-VL-2B, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Only the language model is edited: refusal directions are removed from the text tower's attention output projections (including the convolution operator's out_proj on conv layers) and MLP down-projections. The 164M vision encoder and 25M projector are copied into the GGUF unchanged, so image understanding survives quantization.

[!IMPORTANT] The GGUF is fully multimodal - text and vision. The vision tower ships separately as LFM3-VL-2B-heretic-mmproj-F16.gguf (0.38 GB); the two files together are the model. For image input you must load both - see Quickstart. Text-only use needs only the main GGUF. Nothing is dropped from the text path: the text GGUF carries the full language tower (287 tensors, 24 blocks), and the mmproj carries 189,380,352 vision+projector parameters.

llama.cpp video input for LFM3-VL is not supported upstream yet; image input is tested and works. Video works through the transformers checkpoint.

Who this is for: on-device document, chart, screenshot and video work that should not need a data centre. LFM3-VL-2B is Liquid's post-trained 1.92B vision model, and its published numbers include DocVQA 96.7, TextVQA 87.9, OCRBench v1 86.3, RealWorldQA 79.3 and MMMU 55.7. It is not recommended for coding, multi-turn conversations, or complex reasoning; abliteration removes refusals, not those limits.

Refusals drop from 98/100 to 3/100 at a KL divergence of 0.0014 - a very tight text edit.

Abliteration parameters

Trial 94 of a 200-trial Heretic run (seed 1424362134).

Performance

95 of the 98 refusals clear, and the edit is tight (KL 0.0014). The base was almost fully refusal-tuned on Heretic's harmful evaluation set.

MMLU (text only)

Measured on this checkpoint and on the original model in the same environment (lm-eval 0.4.13, 14,042 questions):

The +0.07 pp difference is far inside one standard error. MMLU exercises the text path only; it does not measure the vision encoder. Vision weights are unchanged from the base model, and the image smoke test in the pipeline checks that the projector still pairs correctly.

Why abliteration instead of fine-tuning

Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.

Files

Safetensors

BF16, 1,918,181,888 parameters (1.92B): 1,728,801,536 language model + 164,206,336 vision encoder + 25,174,016 projector. The reproduce/ directory carries the full Heretic recipe - config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums.

GGUF quantizations

Full quantization set (F16 + 14 quants, 19.31 GB including the projector) produced with llama.cpp master on 2026-10-28.

[!IMPORTANT] The text files use the new lfm3 GGUF architecture and the projector uses clip. You need an LFM3-VL-capable llama.cpp build (support landed in late October 2026); older builds will fail to load them. The upstream LiquidAI/LFM3-VL-2B-GGUF uses the same pair.

Text + vision needs both the chosen text quant and the projector. For reference: Q4_K_M 1.48 GB, Q5_K_M 1.64 GB, Q6_K 1.81 GB, Q8_0 2.23 GB. The projector is not quantized.

Quickstart

# llama.cpp - -hf auto-fetches the matching mmproj (LFM3-VL-capable build)
llama-mtmd-cli -hf saidutta69/LFM3-VL-2B-heretic:Q4_K_M \
  --image photo.jpg -p "Describe this image."

# explicit two-file form
llama-mtmd-cli \
  -m LFM3-VL-2B-heretic-Q4_K_M.gguf \
  --mmproj LFM3-VL-2B-heretic-mmproj-F16.gguf \
  --image photo.jpg -p "Describe this image."

# transformers (the checkpoint ships custom code; needs transformers>=5.21.0)
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "saidutta69/LFM3-VL-2B-heretic"
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "url": "https://placecats.com/300/200"},
        {"type": "text", "text": "Describe this image."},
    ],
}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)
out = model.generate(**inputs, do_sample=True, temperature=0.1, top_p=1.0, top_k=50,
                     repetition_penalty=1.1, max_new_tokens=512)
print(processor.tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Generation parameters

Liquid's recommended settings: temperature=0.1, top_p=1.0, top_k=50, min_p=0.0, repetition_penalty=1.1. For image and video preprocessing use the values in processor_config.json.

Chat template and thinking

LFM3 uses ChatML with <|startoftext|> at the front. Reasoning is enabled by default and arrives in a thinking block; pass enable_thinking=False to disable it (Liquid notes this can be inconsistent on reasoning-heavy tasks). Prior-turn reasoning is dropped unless preserve_thinking=True. Tool calls use the same XML format as LFM3-1.7B; pass schemas with tools=... to apply_chat_template() and feed results back under a tool role.

Runs anywhere

Weights only; the KV cache grows with context. The projector is fixed at 0.38 GB regardless of the text quant.

Model details

Text GGUF: arch=lfm3, 287 tensors, 24 blocks, 128,000 context, tied embeddings (no lm_head tensor). The projector GGUF carries the complete 164,206,336-parameter vision encoder + 25,174,016-parameter projector.

Base-model note: the heretic repo's original YAML tag said LiquidAI/LFM3-1.7B-Base, but the ablation was run on LiquidAI/LFM3-VL-2B (commit f83d2b0a...) - see reproduce/config.toml. This card uses the correct base.

FAQ

What is LFM3-VL-2B-heretic?

LFM3-VL-2B-heretic is a decensored build of LiquidAI/LFM3-VL-2B produced with Heretic 1.4.0 (directional ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.

Is LFM3-VL-2B-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 98/100 to 3/100 at KL divergence 0.0014 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run LFM3-VL-2B-heretic locally?

Text: llama serve -hf saidutta69/LFM3-VL-2B-heretic:Q4_K_M. Images: llama-mtmd-cli -hf saidutta69/LFM3-VL-2B-heretic:Q4_K_M --mmproj LFM3-VL-2B-heretic-mmproj-F16.gguf. The repo ships 15 GGUF quants plus the projector, so vision survives quantization.

Responsible use

Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it should not. There is no safety filtering layered on top. You are responsible for how you deploy it - don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits LFM3-VL-2B's factual and visual limitations; abliteration removes refusal directions, it does not add capability or judgment.

Made with ❤️ by RACER IS OP (saidutta69) - follow for more uncensored models

License

Inherits the LFM Open License v1.0 from the base model. This is not Apache-2.0; review the LFM terms before commercial use.

Related

• LiquidAI/LFM3-VL-2B - the post-trained vision-language base model
• LiquidAI/LFM3-VL-2B-GGUF - the upstream GGUF + mmproj
• LFM3-1.7B-heretic - the text-only sibling
• LFM3 family collection
• RACER IS OP - Heretic Models - full collection

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition (94.97 GB VRAM)
• OS: Linux-4.19.0-gvisor-x86_64-with-glibc2.41
• CUDA / driver: CUDA 13.0, driver 595.71.05
• Software: PyTorch 2.14.1+cu130, Heretic 1.4.0

Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.