LFM2.5-8B-A1B-heretic

A decensored variant of LiquidAI/LFM2.5-8B-A1B, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Liquid's hybrid conv+attention MoE keeps its native reasoning traces, Pythonic tool-call format, and 128K context; 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 and multilingual coverage are left largely intact.

Who this is for: developers who want a fast, uncensored on-device assistant that actually calls tools - the 8.3B-total / 1.5B-active MoE decodes faster than dense models several times its size, so this runs at interactive speed on CPU-only laptops, Apple Silicon, and low-VRAM GPUs while still fitting a 128K context. Pair it with LFM2.5-8B-A1B-DSpark for ~2.5x faster speculative decoding.

Runs on your gaming PC

Full GGUF ladder included - pick the quant that fits your card:

Weights only, at this model's 8.3B total native size; add ~1 GB per 32K of context. Because this is a mixture-of-experts model, memory bandwidth matters far more than raw VRAM - a CPU with fast RAM often outruns a slow GPU. OOM? Drop one quant level. Headroom to spare? Go one up.

Abliteration parameters

Trial 167 of a 200-trial Heretic run (seed 4217170727). direction_index was selected per layer.

Performance

Refusals on the harmful evaluation set drop from 90/100 to 28/100. KL divergence of 0.1649 is the loosest of this batch - the refusal directions in this hybrid conv+attention MoE overlap more with its instruction-following behaviour, so Heretic had to move further to hit the same refusal target. Expect a small amount more drift in style and formatting than a low-KL sibling; the reasoning traces and tool-call format 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, ~8.3B total parameters (~1.5B active per token). 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 (F16 + Q4_K_M, Q5_K_M, Q6_K, Q8_0) produced with llama.cpp.

LFM2 MoE architecture (lfm2moe) - loads natively in llama.cpp / LM Studio / Jan.

Run llama serve -hf saidutta69/LFM2.5-8B-A1B-heretic to pull the default quant.

Quickstart

# llama.cpp
llama serve -hf saidutta69/LFM2.5-8B-A1B-heretic

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

model_name = "saidutta69/LFM2.5-8B-A1B-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "A train leaves at 14:05 travelling 80 km/h. A car leaves at 14:15 travelling 120 km/h on the same track. When does the car catch the train? Show your work."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, do_sample=True, temperature=0.2, top_k=80, repetition_penalty=1.05, max_new_tokens=8192)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

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

Generation parameters

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

model.generate(**inputs, temperature=0.2, top_k=80, repetition_penalty=1.05)

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.

Speculative decoding

Attach LFM2.5-8B-A1B-DSpark (a 328M drafter) for roughly 2.5x faster decoding with identical outputs.

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. It inherits LFM2.5's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.

License

Inherits the LFM License 1.0 from the base model.

Related

• LiquidAI/LFM2.5-8B-A1B — the base model
• LFM2.5 technical report — architecture and benchmarks
• RACER IS OP — Heretic Models — full collection