LFM2.5-350M-heretic

A decensored variant of LiquidAI/LFM2.5-350M, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Liquid's hybrid conv+attention architecture keeps its native Pythonic tool-call format and multilingual coverage; refusal behaviour is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's agentic ability is left largely intact.

Who this is for: developers who want an uncensored assistant that is genuinely fast rather than merely small. LFM2.5-350M decodes at 313 tok/s on an AMD CPU and 188 tok/s on a Snapdragon Gen4 NPU — an order of magnitude faster than a same-size dense Llama — while speaking nine languages. It runs under 1 GB of memory, so it fits in a phone, a Pi, or a browser tab. Reach for it for data extraction, structured output, tool routing, and RAG on hardware where a larger model simply would not run.

This is the most aggressively uncensored model in the collection relative to its size: refusals drop from 89/100 to 4/100, and the 1.2B sibling reaches 2/100.

Runs anywhere

At 350M there is no VRAM ladder — pick a quant by how much RAM and how much bandwidth your device has:

Weights only, at this model's 350M native size. Liquid's own measurements put the whole model — weights and KV cache — under 1 GB at Q4_0, with roughly 32K of usable context. Because the hybrid architecture spends most of its layers in cheap convolution blocks, this model is far more decode-bandwidth-sensitive than its parameter count suggests: fast RAM matters more than a fast GPU.

Abliteration parameters

Trial 176 of a 200-trial Heretic run (seed 3260526957). direction_index was selected per layer.

Performance

Refusals on the harmful evaluation set drop from 89/100 to 4/100. The KL divergence of 0.0989 is the highest in this batch — this base model was heavily refusal-tuned for such a small model, so Heretic had to move further than usual to reach the same target. Expect slightly more drift in style than the low-KL siblings; the tool-call format and multilingual ability survive intact.

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.

Made with ❤️ by RACER IS OP — follow for more uncensored models

Files

Safetensors

BF16, 350M parameters. The reproduce/ directory carries the full Heretic recipe — config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums — so this exact model can be regenerated bit-for-bit. Reproduce it with heretic --reproduce reproduce/reproduce.json.

GGUF quantizations

Full quantization set (14 quants + F16) produced with llama.cpp.

LFM2 dense architecture (lfm2) - loads natively in llama.cpp / LM Studio / Jan. Day-one upstream support, so this also works with MLX, vLLM, and Nexa-optimized NPU builds.

Run llama serve -hf saidutta69/LFM2.5-350M-heretic to pull the default quant.

Quickstart

# llama.cpp
llama serve -hf saidutta69/LFM2.5-350M-heretic

# transformers (requires transformers>=5.0.0)
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/LFM2.5-350M-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "Extract the invoice number and total from: 'Invoice INV-4471, total 231.50 USD, due 30 Nov.'"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, do_sample=True, temperature=0.1, top_k=50, repetition_penalty=1.05, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang, and OpenVINO.

Generation parameters

Liquid's recommended settings - low temperature with a top-k cut:

model.generate(**inputs, temperature=0.1, top_k=50, repetition_penalty=1.05)

Chat template

LFM2.5 uses a ChatML-like format. See the Chat Template documentation for details.

<|startoftext|>system
You are a helpful assistant trained by Liquid AI.
user
What is C. elegans?
assistant

Tool use

LFM2.5 writes Pythonic calls: a Python list between <|tool_call_start|> and <|tool_call_end|>. Pass tool definitions as a JSON list in the system prompt, or hand them to tokenizer.apply_chat_template(..., tools=tools), then feed the result back under a tool role. To get JSON instead, just ask for it in the system prompt.

Model details

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. Liquid explicitly does not recommend LFM2.5-350M for knowledge-intensive tasks or programming; it is built for extraction, structured output, and tool use.

License

Inherits the LFM License 1.0 from the base model.

Related

• LiquidAI/LFM2.5-350M — the base model
• LFM2.5 technical report — architecture and benchmarks
• RACER IS OP — Heretic Models — full collection