LFM3-1.7B-heretic

A decensored variant of LiquidAI/LFM3-1.7B, produced with Heretic v1.4.0 (directional ablation / "abliteration"). LFM3's tri-hybrid architecture is kept intact - 24 layers split into 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers - along with the 256K context window, ChatML template and XML tool-call format. Refusal behaviour is suppressed with targeted weight edits to the attention output projections (including the convolution operator's out_proj on conv layers) and the MLP down-projections, not by fine-tuning.

Who this is for: local agents and tool use on hardware that should not need a GPU. LFM3-1.7B is Liquid's post-trained 1.7B agentic model, and its published LFM3 column reports BFCL V4 71.39, ToolSandbox 80.36 and WebArena-Verified 38.42 - numbers that used to need a 9B model. At 4-bit the weights are about 1 GB, so it fits a phone, a Pi 5, or a browser tab. Liquid itself does not recommend it for agentic coding or complex multi-step reasoning at this size; the abliteration removes refusals, not that limitation.

Refusals drop from 95/100 to 5/100 at a KL divergence of 0.0052, and MMLU is unchanged within noise (49.91% vs 49.56%).

Abliteration parameters

Trial 83 of a 200-trial Heretic run (seed 2678800539).

Performance

The base model was strongly refusal-tuned (95/100 on Heretic's harmful evaluation set), and 90 of those 95 clear. KL 0.0052 is a tight edit, so expect little drift in phrasing or format.

MMLU

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

The +0.35 pp difference is smaller than one standard error, i.e. no measurable knowledge regression on MMLU. It is a single benchmark, not a capability guarantee.

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, 1,728,801,536 parameters (1.73B). 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. Reproduce it with heretic --reproduce reproduce/reproduce.json.

GGUF quantizations

Full quantization set (F16 + 14 quants, 18.93 GB total) produced with llama.cpp master on 2026-10-28.

[!IMPORTANT] These files use the new lfm3 GGUF architecture. You need an LFM3-capable llama.cpp build (support landed in late October 2026); older builds will fail to load the model. The upstream LiquidAI/LFM3-1.7B-GGUF uses the same architecture.

Quickstart

# llama.cpp (LFM3-capable build)
llama serve -hf saidutta69/LFM3-1.7B-heretic

# transformers (the checkpoint ships custom code)
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/LFM3-1.7B-heretic"
model = AutoModelForCausalLM.from_pretrained(
    model_name, device_map="auto", dtype="bfloat16", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "Explain how a heat pump works."}]
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=1.1, max_new_tokens=2048,
)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Generation parameters

Liquid's recommended settings:

• General (chat, tool use, agents): temperature=0.1, top_k=50, repetition_penalty=1.1
• Math & STEM: temperature=1.0, top_p=0.95, top_k=20

Chat template and thinking

LFM3 uses ChatML: <|im_start|>role ... <|im_end|> 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 use

Pass function schemas with tools=... to tokenizer.apply_chat_template(). Calls come back as XML:

<tool_call>
<function=get_weather>
<parameter=city>
Paris
</parameter>
</function>
</tool_call>

Feed the result back under a tool role. Deferred tool loading (defer_loading=True) and the <additional_tools> search flow are supported by the template.

Runs anywhere

Weights only. The hybrid stack spends 13 of its 24 layers in cheap convolution blocks, so decode is memory-bandwidth-bound rather than compute-bound. Liquid measures about 1 GB of weights at 4-bit; the KV cache grows with context, and at the full 256K window it is no longer a phone workload.

Model details

The GGUF carries 287 tensors over 24 blocks (arch=lfm3) and exactly 262,144 context. Embeddings are tied, so there is no separate lm_head tensor.

Base-model note: the heretic repo's original YAML tag said LiquidAI/LFM3-1.7B-Base, but the ablation was run on the post-trained LiquidAI/LFM3-1.7B (commit 40ebe0aa...) - see reproduce/config.toml. This card uses the correct base.

FAQ

What is LFM3-1.7B-heretic?

LFM3-1.7B-heretic is a decensored build of LiquidAI/LFM3-1.7B 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 LFM3-1.7B-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 95/100 to 5/100 at KL divergence 0.0052 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run LFM3-1.7B-heretic locally?

llama serve -hf saidutta69/LFM3-1.7B-heretic:Q4_K_M pulls the default quant; the repo ships 15 GGUF quants (see Files).

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 LFM3-1.7B's factual 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/LFM3-1.7B - the post-trained base model
• LiquidAI/LFM3-1.7B-Base - the pretrained checkpoint
• LFM3 family collection - base, text, vision, audio, omni, decision
• 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.