LFM2.5-2.6B-heretic

A decensored variant of LiquidAI/LFM2.5-2.6B, produced with Heretic 2.0.0.dev0 using Arbitrary-Rank Ablation (ARA). LFM2.5-2.6B is Liquid's on-device agentic model: 30 layers laid out as 22 double-gated short convolution blocks plus 8 GQA layers, a 131K context, and a native tool-call format. ARA edits the language model over layers 1-21; everything else is unchanged.

Who this is for: local agents, tool use, RAG and structured extraction where a 2.7B model should run on a phone or a small laptop. Liquid's published numbers include BFCLv4 56.88, ToolSandbox 77.83, IFBench 59.17 and IFStruct 85.49, and it is not recommended for agentic coding or knowledge-heavy work at this size. This checkpoint is not the Fable5 coding-agent fine-tune (lfm2.5-2.6b-fable5-coding-agent-heretic); it is the base 2.6B instruct model with refusals removed.

Refusals drop from 96/100 to 8/100 at a KL divergence of 0.0059 - a tight edit.

Arbitrary-Rank Ablation parameters

Trial 190 of a 250-trial Heretic 2.0 ARA run (seed 712709786).

Performance

88 of the 96 refusals clear. KL 0.0059 is a tight edit, so expect little drift in phrasing, tool format or multilingual behaviour.

Base model benchmarks

Liquid's published numbers for LFM2.5-2.6B, not re-run here:

Abliteration removes refusal directions; it does not add capability. Treat these as the base model's numbers, not a re-measurement of this checkpoint.

Why ARA instead of directional ablation

Directional ablation finds one refusal direction and orthogonalizes attention/MLP outputs against it. ARA learns a low-rank subspace over a layer range and blends a "preserve good behaviour" and "steer bad behaviour" objective. Here it kept 96% of the preserve objective over layers 1-21 and still removed 88 of 96 refusals. See the Heretic repo.

Files

Safetensors

BF16, 2,697,198,592 parameters (2.69B). The reproduce/ directory carries the full Heretic 2.0 recipe - config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums.

GGUF quantizations

Full quantization set (F16 + 14 quants, 28.94 GB total), produced with llama.cpp master on 2026-10-28.

Text GGUF: arch lfm2, 266 tensors, 30 blocks (22 convolution + 8 full attention), 131,072 context, 128,000 vocab, tied embeddings, conv cache L_cache=3, rope theta 1e6.

Quickstart

# llama.cpp
llama serve -hf saidutta69/LFM2.5-2.6B-heretic

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/LFM2.5-2.6B-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", dtype="bfloat16")
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=1024)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Generation parameters

Liquid's recommended settings: temperature=0.1, top_k=50, repetition_penalty=1.1.

Chat template and tools

LFM2.5 uses ChatML-style tags and Liquid's native tool-call format. Pass tool schemas with tools=... to apply_chat_template() and feed results back under a tool role; see the base card for the exact format.

Runs anywhere

Weights only. Liquid measures about 30 tok/s on a phone with the 2.6B model, and the hybrid stack spends 22 of its 30 layers in cheap convolution blocks, so decode is bandwidth-bound.

Model details

FAQ

What is LFM2.5-2.6B-heretic?

LFM2.5-2.6B-heretic is a decensored build of LiquidAI/LFM2.5-2.6B produced with Heretic 2.0.0.dev0 (Arbitrary-Rank Ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.

Is LFM2.5-2.6B-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 96/100 to 8/100 at KL divergence 0.0059 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run LFM2.5-2.6B-heretic locally?

llama serve -hf saidutta69/LFM2.5-2.6B-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 LFM2.5-2.6B'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/LFM2.5-2.6B - the base agentic model
• LFM2.5 collection
• lfm2.5-2.6b-fable5-coding-agent-heretic - the separate Fable5 coding-agent fine-tune
• 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 2.0.0.dev0

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