This is a decensored version of LiquidAI/LFM2.5-1.2B-Instruct, made using Heretic v2.0.0.dev0

[!TIP] This model is reproducible!

See the README in the reproduce directory for more information.

Abliteration parameters

Performance

Summary

Abliterated build of LiquidAI/LFM2.5-1.2B-Instruct, Liquid AI's 1.2B edge model with a hybrid convolution + attention architecture, built for fast on-device inference. GGUF quantizations: richardyoung/LFM2.5-1.2B-Instruct-heretic-GGUF · Ollama: richardyoung/lfm2.5-1.2b-instruct-heretic

Abliteration Results

The very low KL divergence (0.059, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

Quick Start

With Ollama

ollama run richardyoung/lfm2.5-1.2b-instruct-heretic
# or straight from Hugging Face
ollama run hf.co/richardyoung/LFM2.5-1.2B-Instruct-heretic-GGUF

With llama.cpp

huggingface-cli download richardyoung/LFM2.5-1.2B-Instruct-heretic-GGUF \
    --include "*Q4_K_M*" --local-dir ./models

./llama-cli -m ./models/*Q4_K_M*.gguf -ngl 99

With Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="richardyoung/LFM2.5-1.2B-Instruct-heretic-GGUF",
    filename="*Q4_K_M*",
    n_gpu_layers=-1,
)

output = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Explain abliteration in simple terms."}]
)
print(output["choices"][0]["message"]["content"])

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "richardyoung/LFM2.5-1.2B-Instruct-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")
messages = [{"role": "user", "content": "Explain abliteration in simple terms."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(inputs, max_new_tokens=256)[0][inputs.shape[-1]:], skip_special_tokens=True))

What is Abliteration?

Abliteration removes the "refusal direction" from a model's residual stream — a surgical modification that disables safety refusals without retraining. See Refusal in Language Models Is Mediated by a Single Direction. Heretic automates it: it searches ablation parameters with an optimizer (200 trials here) and keeps the trials that best trade off fewer refusals against KL divergence from the original model.

Intended Use

Research, creative writing, education on alignment techniques, and unrestricted local inference. This model will comply with requests the original refuses; use responsibly and in accordance with applicable laws.

Other Models by richardyoung

• New Heretic models (reproducible, listed in the Heretic Grimoire): LFM2.5-1.2B | Granite-4.2-3B | Granite-4.2-8B | Olmo-3-7B-Think | Phi-4-mini | DeepSeek-R1-0528-Qwen3-8B | Agents-A1-4B
• Abliterated/Uncensored models: Qwen2.5-7B | Qwen3-14B | DeepSeek-R1-32B | Qwen3-8B
• OCR & Vision: MiniCPM-V-4.6 GGUF | dots.mocr GGUF | Unlimited-OCR GGUF | olmOCR GGUF
• MLX quantizations (Apple Silicon): Kimi-K2 series | olmOCR MLX
• Healthcare/Medical: Synthea 575K patients dataset | CardioEmbed
• Research: LLM Instruction-Following Evaluation (arxiv:2510.18892)

Built & maintained by Richard Young · DeepNeuro

The original model card follows below.

LFM2.5-1.2B-Instruct

LFM2.5 is a new family of hybrid models designed for on-device deployment. It builds on the LFM2 architecture with extended pre-training and reinforcement learning.

• Best-in-class performance: A 1.2B model rivaling much larger models, bringing high-quality AI to your pocket.
• Fast edge inference: 239 tok/s decode on AMD CPU, 82 tok/s on mobile NPU. Runs under 1GB of memory with day-one support for llama.cpp, MLX, and vLLM.
• Scaled training: Extended pre-training from 10T to 28T tokens and large-scale multi-stage reinforcement learning.

Find more information about LFM2.5 in our blog post.

🗒️ Model Details

LFM2.5-1.2B-Instruct is a general-purpose text-only model with the following features:

• Number of parameters: 1.17B
• Number of layers: 16 (10 double-gated convolution blocks + 6 GQA blocks)
• Training budget: 28T tokens
• Context length: 32,768 tokens
• Vocabulary size: 65,536
• Knowledge cutoff: Mid-2024
• Languages: English, Arabic, Chinese, French, German, Japanese, Korean, Spanish
• Generation parameters:
  • temperature: 0.1
  • top_k: 50
  • repetition_penalty: 1.05

We recommend using it for agentic tasks, data extraction, and RAG. It is not recommended for knowledge-intensive tasks and programming.

Chat Template

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

<|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

You can use tokenizer.apply_chat_template() to format your messages automatically.

Tool Use

LFM2.5 supports function calling as follows:

1. Function definition: We recommend providing the list of tools as a JSON object in the system prompt. You can also use the tokenizer.apply_chat_template() function with tools.
2. Function call: By default, LFM2.5 writes Pythonic function calls (a Python list between <|tool_call_start|> and <|tool_call_end|> special tokens), as the assistant answer. You can override this behavior by asking the model to output JSON function calls in the system prompt.
3. Function execution: The function call is executed, and the result is returned as a "tool" role.
4. Final answer: LFM2 interprets the outcome of the function call to address the original user prompt in plain text.

See the Tool Use documentation for the full guide. Example:

<|startoftext|><|im_start|>system
List of tools: [{"name": "get_candidate_status", "description": "Retrieves the current status of a candidate in the recruitment process", "parameters": {"type": "object", "properties": {"candidate_id": {"type": "string", "description": "Unique identifier for the candidate"}}, "required": ["candidate_id"]}}]<|im_end|>
<|im_start|>user
What is the current status of candidate ID 12345?<|im_end|>
<|im_start|>assistant
<|tool_call_start|>[get_candidate_status(candidate_id="12345")]<|tool_call_end|>Checking the current status of candidate ID 12345.<|im_end|>
<|im_start|>tool
[{"candidate_id": "12345", "status": "Interview Scheduled", "position": "Clinical Research Associate", "date": "2023-11-20"}]<|im_end|>
<|im_start|>assistant
The candidate with ID 12345 is currently in the "Interview Scheduled" stage for the position of Clinical Research Associate, with an interview date set for 2023-11-20.<|im_end|>

🏃 Inference

LFM2.5 is supported by many inference frameworks. See the Inference documentation for the full list.

[!TIP] ⚡ Faster decoding: attach LFM2.5-1.2B-Instruct-DSpark, a 296M speculative-decoding drafter, for ~2.1x faster decoding in SGLang and ~2.5x on Apple silicon via Metal with exactly the same outputs.

Here's a quick start example with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "LiquidAI/LFM2.5-1.2B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16",
#   attn_implementation="flash_attention_2" <- uncomment on compatible GPU
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

prompt = "What is C. elegans?"

input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.05,
    max_new_tokens=512,
    streamer=streamer,
)

🔧 Fine-Tuning

We recommend fine-tuning LFM2.5 for your specific use case to achieve the best results.

📊 Performance

Benchmarks

We compared LFM2.5-1.2B-Instruct with relevant sub-2B models on a diverse suite of benchmarks.

GPQA, MMLU-Pro, IFBench, and AIME25 follow ArtificialAnalysis's methodology. For IFEval and Multi-IF, we report the average score across strict and loose prompt and instruction accuracies. For BFCLv3, we report the final weighted average score with a custom Liquid handler to support our tool use template.

Inference speed

LFM2.5-1.2B-Instruct offers extremely fast inference speed on CPUs with a low memory profile compared to similar-sized models.

In addition, we are partnering with AMD, Qualcomm, and Nexa AI to bring the LFM2.5 family to NPUs. These optimized models are available through our partners, enabling highly efficient on-device inference. The following numbers have been calculated using 1K prefill and 100 decode tokens:

These capabilities unlock new deployment scenarios across various devices, including vehicles, mobile devices, laptops, IoT devices, and embedded systems.

📬 Contact

• Got questions or want to connect? Join our Discord community
• If you are interested in custom solutions with edge deployment, please contact our sales team.

Citation

@article{liquidai2025lfm2,
  title={LFM2 Technical Report},
  author={Liquid AI},
  journal={arXiv preprint arXiv:2511.23404},
  year={2025}
}