LFM2.5-1.2B-Thinking-heretic

RACER IS OP

A decensored variant of LiquidAI/LFM2.5-1.2B-Thinking, produced with Heretic v1.4.0 (directional ablation / "abliteration"). This is the reasoning sibling of the LFM2.5-1.2B line: it emits an explicit <think>…</think> trace before its answer. Liquid's hybrid conv+attention architecture keeps its native Pythonic tool-call format, instruction-following, 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 reasoning ability is left largely intact.

Who this is for: developers who want a genuine on-device reasoning model that still fits on a phone. LFM2.5-1.2B-Thinking scores 87.96 on MATH-500, 85.60 on GSM8K, 88.42 on IFEval and 56.97 on BFCLv3 β€” real chain-of-thought and tool-calling numbers β€” while decoding at 116 tok/s on an AMD Ryzen CPU (Q4_0) in under 1 GB. Use it for local agents, math and multi-step reasoning, RAG, and structured extraction where a cloud round-trip would be the bottleneck.

Refusals drop from 98/100 to 10/100. See Performance for why the KL figure here needs reading together with that count.

Runs anywhere

Device Recommended quant Weights
Phone / browser (4 GB RAM) Q4_K_M ~0.68 GB
Raspberry Pi 5 (4 GB) Q4_K_M ~0.68 GB
Laptop (8 GB, Apple Silicon) Q8_0 ~1.16 GB
RTX 3060 / 4060 (8-12 GB) Q6_K ~0.90 GB
Server / multi-tenant Q5_K_M ~0.79 GB

Weights only, at this model's 1.17B native size. The hybrid architecture spends 10 of its 16 layers in cheap convolution blocks, so this model is decode-bandwidth-bound: fast RAM and memory bandwidth matter far more than raw compute. Because it is a reasoning model, budget output tokens generously β€” the <think> trace can dwarf the final answer on a hard question. OOM? Drop one quant level; the Q2_K tier (~0.45 GB) is there if you truly need it.

Abliteration parameters

Trial 83 of a 200-trial Heretic run (seed 2205080409).

Parameter Value
direction_index per layer
attn.o_proj.max_weight 1.49
attn.o_proj.max_weight_position 11.43
attn.o_proj.min_weight 1.48
attn.o_proj.min_weight_distance 7.75
mlp.down_proj.max_weight 1.44
mlp.down_proj.max_weight_position 9.61
mlp.down_proj.min_weight 1.12
mlp.down_proj.min_weight_distance 8.36

Performance

Metric This model Original model (LiquidAI/LFM2.5-1.2B-Thinking)
KL divergence 0.0005 0 (by definition)
Refusals 10/100 98/100

Refusals on the harmful evaluation set drop from 98/100 to 10/100. The KL divergence of 0.0005 is the lowest in this collection: the optimiser found a direction it applied conservatively, which is why so much of the base model's behaviour and reasoning survives. It is not a free lunch β€” 10 refusals remain, more than the near-fully-decensored edits in this collection. Read the KL figure and the refusal count together, never the KL alone: a low KL only means the edit was careful, not that refusals fell.

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 (saidutta69) β€” follow for more uncensored models

Files

Safetensors

File Size
model.safetensors 2.18 GB

BF16, 1.17B parameters. The reproduce/ directory carries the full Heretic recipe β€” config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums β€” 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.

File Format Size
LFM2.5-1.2B-Thinking-heretic-F16.gguf GGUF F16 2.18 GB
LFM2.5-1.2B-Thinking-heretic-Q2_K.gguf GGUF Q2_K 0.45 GB
LFM2.5-1.2B-Thinking-heretic-IQ3_S.gguf GGUF IQ3_S 0.52 GB
LFM2.5-1.2B-Thinking-heretic-Q3_K_S.gguf GGUF Q3_K_S 0.52 GB
LFM2.5-1.2B-Thinking-heretic-Q3_K_M.gguf GGUF Q3_K_M 0.56 GB
LFM2.5-1.2B-Thinking-heretic-Q3_K_L.gguf GGUF Q3_K_L 0.59 GB
LFM2.5-1.2B-Thinking-heretic-IQ4_XS.gguf GGUF IQ4_XS 0.62 GB
LFM2.5-1.2B-Thinking-heretic-Q4_K_S.gguf GGUF Q4_K_S 0.65 GB
LFM2.5-1.2B-Thinking-heretic-Q4_0.gguf GGUF Q4_0 0.65 GB
LFM2.5-1.2B-Thinking-heretic-Q4_1.gguf GGUF Q4_1 0.71 GB
LFM2.5-1.2B-Thinking-heretic-Q4_K_M.gguf GGUF Q4_K_M 0.68 GB
LFM2.5-1.2B-Thinking-heretic-Q5_K_S.gguf GGUF Q5_K_S 0.77 GB
LFM2.5-1.2B-Thinking-heretic-Q5_K_M.gguf GGUF Q5_K_M 0.79 GB
LFM2.5-1.2B-Thinking-heretic-Q6_K.gguf GGUF Q6_K 0.90 GB
LFM2.5-1.2B-Thinking-heretic-Q8_0.gguf GGUF Q8_0 1.16 GB

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

Run llama serve -hf saidutta69/LFM2.5-1.2B-Thinking-heretic to pull the default quant.

Quickstart

# llama.cpp
llama serve -hf saidutta69/LFM2.5-1.2B-Thinking-heretic
# transformers (requires transformers>=5.0.0)
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/LFM2.5-1.2B-Thinking-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=0.05, top_k=50, max_new_tokens=4096)
text = tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)

# The model emits a <think>...</think> trace, then the answer. Strip it for display,
# but keep it in the message history on multi-turn so the model stays consistent.
answer = text.split("</think>")[-1].strip() if "</think>" in text else text
print(answer)

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

Reasoning format

This is a thinking model. It writes its chain of thought between <think> and </think> before the final answer. The chat template keeps the trace for the most recent assistant turn by default; pass preserve_thinking=True to apply_chat_template to keep it across turns. Note that output can contain a closing </think> with the opening tag supplied by the template, so strip on </think> rather than assuming both tags are present.

Generation parameters

Liquid's recommended settings for this reasoning model - a very low temperature with a top-k cut:

model.generate(**inputs, temperature=0.05, top_k=50)

Chat template

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

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

Architecture Lfm2ForCausalLM (hybrid conv + attention)
Parameters 1.17B
Layers / heads 16 layers (10 double-gated conv + 6 GQA), 32 attention heads, 8 KV heads
Hidden / intermediate 2048 / 12288
Convolution cache L_cache = 3
Position embedding RoPE, theta = 1,000,000
Context length 32,768 tokens
Vocab 65,536
Precision bfloat16
Languages English, Arabic, Chinese, French, German, Japanese, Korean, Spanish
Base model LiquidAI/LFM2.5-1.2B-Thinking

FAQ

What is LFM2.5-1.2B-Thinking-heretic?

LFM2.5-1.2B-Thinking-heretic is a decensored build of LiquidAI/LFM2.5-1.2B-Thinking 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 LFM2.5-1.2B-Thinking-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 98/100 to 10/100 at KL divergence 0.0005 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run LFM2.5-1.2B-Thinking-heretic locally?

llama serve -hf saidutta69/LFM2.5-1.2B-Thinking-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 shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it β€” don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits LFM2.5's factual limitations; abliteration removes refusal directions, it doesn't add capability or judgment.

License

Inherits the LFM License 1.0 from the base model.

Related

Abliteration hardware

Recorded in reproduce/reproduce.json:

  • GPU: NVIDIA GeForce RTX 4090 (23.52 GB VRAM)
  • CPU: AMD EPYC 7402 24-Core Processor
  • OS: Linux-6.8.0-146-generic-x86_64-with-glibc2.35
  • CUDA / driver: CUDA 13.0, driver 595.99.02
  • 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.

Downloads last month
1,409
Safetensors
Model size
1B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for saidutta69/LFM2.5-1.2B-Thinking-heretic

Quantized
(58)
this model

Collection including saidutta69/LFM2.5-1.2B-Thinking-heretic

Paper for saidutta69/LFM2.5-1.2B-Thinking-heretic