LFM2.5-VL-3B-heretic

A decensored variant of LiquidAI/LFM2.5-VL-3B, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Liquid's hybrid conv+attention text backbone and the SigLIP2 NaFlex vision tower are kept intact - refusal behaviour is suppressed with targeted weight edits to the attention output and MLP down-projections of the language model, not by fine-tuning. Visual grounding, full-page OCR with layout annotation, multi-image input and tool calling all still work.

[!IMPORTANT] The GGUF here is fully multimodal - text and vision. The vision tower and projector ship separately as LFM2.5-VL-3B-heretic-mmproj-F16.gguf (0.85 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 (266 tensors, 30 blocks), and the projector GGUF carries the complete 426,285,296-element SigLIP2 tower plus merger, measured from its own header, including the tiling metadata (min_tiles 2, max_tiles 10, tile_size 512, max_pixels_tolerance 2.0) llama.cpp needs to reproduce the original image preprocessing.

Who this is for: people who want a small vision-language model that answers instead of declining. At 3.12B total (2.70B text + 0.41B SigLIP2 vision + 0.01B projector) it is the larger of the two LFM2.5-VL heretics here, and it is built for document and screen work: full-page OCR with layout regions, bounding-box grounding from a natural-language query, screenshots and UI understanding, chart reading, and multi- image input. Refusals drop from 100/100 to 6/100. 16 languages. If you want the same capability at a fraction of the size, LFM2.5-VL-450M-heretic is the sibling; if you need text only, LFM2.5-1.2B-Instruct-heretic or Qwen3-4B-Thinking-heretic fit better.

Not the right pick for long-context or reasoning-heavy work. Liquid explicitly recommends this family for single-turn, high-throughput, low-latency tasks and warns against visual web design or highly technical blueprint Q&A. Context is 32,768 tokens.

GPU matrix

The vision tower is not quantized, so image input costs the text quant plus the fixed 0.85 GB projector at every tier:

Add roughly 1-2 GB of KV/activation overhead on top. Liquid reports about 3.3 GB resident for the unquantized model, so the Q4_K_M pair comfortably fits a 6 GB card.

Abliteration parameters

Trial 67 of 200 (seed 3853387823), direction_index selected per layer.

Performance

Read the pair, not the KL alone. 0.0645 is a mid-pack edit: looser than granite-4.2-3b (0.0095) and Qwen3-4B-Thinking (0.0011), tighter than granite-4.1-3b (0.0852) and the gemma-4-E2B run (0.1841). The base refused all 100 evaluation prompts - heavier alignment than the Qwen2.5-VL-3B base (99/100) - and 94 of them now clear. That is real work at a weight distance that did not have to be extreme, which is the combination worth trusting.

The base model's own published scores (the abliteration only removes refusal directions, so these describe the underlying capability, not a fresh sweep of this checkpoint):

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 - including the vision tower, OCR behaviour and tool-call format - untouched. See the Heretic repo and the original abliteration writeup for the mechanism.

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

Files

Safetensors

BF16, 3.12B total - 2.70B text tower + 0.41B SigLIP2 vision tower + 0.01B multimodal projector (6.25 GB on disk). The reproduce/ directory carries the full Heretic recipe - config.toml, requirements.txt, the Optuna journal, SHA-256 sums - so this exact model can be regenerated with heretic --reproduce reproduce/reproduce.json.

GGUF quantizations

Full ladder (14 quants + F16) from llama.cpp, plus the vision projector:

Total GGUF payload: 29.79 GB across 16 files (28.94 GB of text weights + 0.85 GB projector).

The text tower is GGUF architecture lfm2 (266 tensors, 30 blocks - 22 convolution and 8 full-attention, at indices 2, 5, 9, 13, 17, 21, 24 and 27), context 32,768. The projector is GGUF architecture clip with projector type lfm2 (441 tensors, the 27-block SigLIP2-so400m tower at hidden 1,152, patch 16, projection dim 2,048). Both load natively in llama.cpp, LM Studio, Jan and Ollama.

Quickstart

Text only

# pulls the default quant for you
llama-completion -hf saidutta69/LFM2.5-VL-3B-heretic:Q4_K_M -ngl 99

Text and images

llama-mtmd-cli auto-discovers the sibling whose name contains mmproj:

llama-mtmd-cli \
  -hf saidutta69/LFM2.5-VL-3B-heretic:Q4_K_M \
  --image path/to/page.png \
  -p "Transcribe this page." \
  -ngl 99

# or pass the projector explicitly from a local checkout
llama-mtmd-cli \
  -m LFM2.5-VL-3B-heretic-Q4_K_M.gguf \
  --mmproj LFM2.5-VL-3B-heretic-mmproj-F16.gguf \
  --image page.png \
  -p "Transcribe this page."

# transformers (vision-language), needs transformers>=5.0.0
from transformers import AutoProcessor, AutoModelForImageTextToText

model_name = "saidutta69/LFM2.5-VL-3B-heretic"
model = AutoModelForImageTextToText.from_pretrained(model_name, dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_name)

conversation = [
    {"role": "user", "content": [
        {"type": "image", "url": "https://example.com/photo.jpg"},
        {"type": "text", "text": "Describe this image in one sentence."},
    ]},
]
inputs = processor.apply_chat_template(
    conversation, add_generation_prompt=True, return_tensors="pt",
    return_dict=True, tokenize=True,
).to(model.device)
outputs = model.generate(**inputs, do_sample=True, temperature=0.2, top_k=50,
                         repetition_penalty=1.0, max_new_tokens=256)
print(processor.batch_decode(outputs[:, inputs["input_ids"].shape[1]:],
                             skip_special_tokens=True)[0])

config.json uses dtype: rather than torch_dtype:, so use dtype="auto" and a recent transformers. The reproduce/requirements.txt pins 5.20.2 for exact regeneration.

Visual grounding

Ask in words; it answers with normalized boxes:

prompt = ('Detect all instances of: the red button. Response must be a JSON array '
          '[{"label": ..., "bbox": [x1, y1, x2, y2]}, ...]. '
          'Coordinates are normalized to [0,1].')

OCR with layout

The model can return reading-order layout regions (image_index=<n> <label> [xmin, ymin, xmax, ymax] followed by the region content, with OTSL for tables and LaTeX for equations). Ask for the format explicitly - it is documented but still experimental upstream.

Sampling

Text defaults from the base card: temperature=0.2, top_k=50, repetition_penalty=1.0. Vision preprocessing is read from processor_config.json; do not override it in llama.cpp, the projector GGUF already carries the tiling parameters.

Model details

About LFM2.5-VL

LFM2.5-VL is Liquid AI's refresh of LFM2-VL, built on the LFM2.5-2.6B language model. It keeps SigLIP2 NaFlex native-resolution processing (large images split into 512x512 tiles with a thumbnail), adds full-page OCR with layout annotation, improves grounding, and still exposes Pythonic tool calls. Liquid quotes 228 tok/s on an Apple M5 Max, 116 tok/s on an AMD Ryzen AI Max+ 395, about 3.3 GB resident, and 20 tok/s on a Galaxy S26 Ultra. It is tuned for single-turn, high-throughput, low-latency work rather than long-context reasoning.

FAQ

What is LFM2.5-VL-3B-heretic?

LFM2.5-VL-3B-heretic is a decensored build of LiquidAI/LFM2.5-VL-3B 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 LFM2.5-VL-3B-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 100/100 to 6/100 at KL divergence 0.0645 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run LFM2.5-VL-3B-heretic locally?

Text: llama serve -hf saidutta69/LFM2.5-VL-3B-heretic:Q4_K_M. Images: llama-mtmd-cli -hf saidutta69/LFM2.5-VL-3B-heretic:Q4_K_M --mmproj LFM2.5-VL-3B-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 shouldn't. 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 the base model's factual limitations and biases; abliteration removes refusal directions, it does not add capability or judgment.

License

lfm1.0 - the LFM Open License, inherited from the base model. This is not Apache-2.0. The heretic repository itself ships no LICENSE file, so the link above points at the base repository's license blob. Read it before deploying commercially.

Related

• LiquidAI/LFM2.5-VL-3B - the base model

• LiquidAI/LFM2.5-VL-3B-GGUF - Liquid's own GGUF, for comparison

• saidutta69/LFM2.5-VL-450M-heretic - the tiny multimodal sibling

• saidutta69/LFM2.5-1.2B-Instruct-heretic - text-only LFM2.5

• RACER IS OP - Heretic Models - full collection - every heretic model in this 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.