LFM2.5-1.2B-Thinking-Heretic

LFM2.5-1.2B-Thinking Heretic

An experimental abliterated derivative of LiquidAI/LFM2.5-1.2B-Thinking. It attenuates an internal direction associated with refusal while attempting to preserve the base model's behavior on benign prompts.

Best measured LFM2.5-1.2B-Thinking Heretic checkpoint in our apples-to-apples comparison. It matched the best refusal-marker score of the public Heretic reference point (3/100) while reducing benign-prompt KL divergence from 0.0007 to 0.0003. This is the strongest checkpoint we measured with this reproducible harness at publication time; it is not a universal ranking of unevaluated Hugging Face derivatives.

This repository contains two non-quantized BF16 variants:

• model.safetensors: native Transformers checkpoint;
• LFM2.5-1.2B-Thinking-heretic-BF16.gguf: GGUF BF16 for llama.cpp and LM Studio.

Modified model notice: the weights differ from the Liquid AI checkpoint. This derivative was produced independently using the Heretic abliteration method. It is not created, endorsed, or supported by Liquid AI.

Results

The selected Optuna trial was reloaded from disk and evaluated separately.

Evaluation setup: 100 refusal-oriented prompts, 100 benign prompts, BF16, seed 42, and 60 Optuna trials. The refusal metric is a lexical proxy, not a complete safety evaluation. KL divergence only covers the configured benign prompt set. Reasoning quality, multilingual quality, factuality, instruction following, and safety require broader evaluation.

Both the original and modified checkpoints can spend more than 512 tokens in <think> on very simple instructions. This behavior was observed in the base checkpoint and is not unique to this derivative.

Method

Per-layer refusal directions were estimated from residual activations on two prompt groups. Output weights were projected away from these directions while preserving row norms. Heretic was adapted for LFM2.5's hybrid architecture: 10 convolutional LIV blocks and 6 GQA blocks. Optuna searched separate strength profiles for convolution, attention, and MLP output projections.

Pinned sources:

• Base model revision: f313478934a7612d22991f752959d7a1a8756fec
• Heretic revision: bedb94ef117a271532ac2058447fbc165d5051bd

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "0xzknw/LFM2.5-1.2B-Thinking-Heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="bfloat16",
    device_map="auto",
)

messages = [{"role": "user", "content": "Explain why the sky is blue."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Recommended base-model generation settings: temperature=0.05, top_k=50, and repetition_penalty=1.05.

LM Studio / llama.cpp

Download LFM2.5-1.2B-Thinking-heretic-BF16.gguf and open it in LM Studio. This GGUF stores the model in BF16; it is a container conversion, not a reduced bit-width quantization. It is approximately 2.34 GB.

llama-cli -m LFM2.5-1.2B-Thinking-heretic-BF16.gguf -cnv

The GGUF was converted with llama.cpp tag b10603, commit c060ca974c773c7c3d17fd1b66dc9d312bc292c0, then loaded and tested in LM Studio. A local smoke test produced approximately 95 tokens/s on the test machine; throughput varies by hardware and runtime settings.

Checksums

Safety and limitations

Abliteration deliberately weakens refusal behavior. This may remove useful safeguards, increase harmful compliance, or cause unexpected regressions. Do not treat this checkpoint as safety-aligned. Evaluate it for your use case, keep human oversight for consequential applications, and follow applicable laws and platform policies.

The model inherits the capabilities, limitations, supported languages, and knowledge cutoff of the base model except where the weight modification changes them. See the original model card for architecture and base-model details.

License and attribution

Distributed under the LFM Open License v1.0, inherited from the base model. The license includes a commercial-use revenue threshold; read LICENSE before use or redistribution. Original model copyright and attribution belong to Liquid AI, Inc. Modifications and evaluation described here were performed by the repository author.

References

• Liquid AI, LFM2.5-1.2B-Thinking: On-Device Reasoning Under 1GB (2026).
• Arditi et al., Refusal in Language Models Is Mediated by a Single Direction (2024), arXiv:2406.11717.
• p-e-w/heretic.