LFM2.5-8B-A1B

LFM2.5-8B-A1B

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.

• On-device personal assistant: Designed to power real-life applications, chaining tool calls, and following complex instructions on all devices.
• Compressed performance: Competitive with much larger dense and MoE models on instruction following and agentic tasks.
• Unmatched throughput: Fastest in its size class on both CPU and GPU inference, with day-one support for llama.cpp, MLX, vLLM, and SGLang.

Find more information about LFM2.5-8B-A1B in our blog post.

*AA-Omniscience Index (higher is better) rewards correct answers and penalizes hallucinations. Scores range from -100 to 100. See more results on Artificial Analysis.

🗒️ Model Details

Model	Parameters	Description
LFM2.5-8B-A1B-Base	8.3B total / 1.5B active	Pre-trained base model for fine-tuning
LFM2.5-8B-A1B	8.3B total / 1.5B active	Reasoning-tuned general-purpose model

LFM2.5-8B-A1B is a general-purpose text-only model with the following features:

• Total parameters: 8.3B
• Active parameters: 1.5B
• Number of layers: 24 (18 double-gated conv + 6 GQA)
• Training budget: 38 trillion tokens
• Context length: 128,000
• Vocabulary size: 128,000
• Languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish
• Generation parameters: We recommend the following parameters:
  • temperature: 0.2
  • top_k: 80
  • repetition_penalty: 1.05

Model	Description
LFM2.5-8B-A1B	Original model checkpoint in native format. Best for fine-tuning or inference with Transformers, vLLM, and SGLang.
LFM2.5-8B-A1B-GGUF	Quantized format for llama.cpp and compatible tools. Optimized for edge inference and local deployment.
LFM2.5-8B-A1B-ONNX	ONNX Runtime format for cross-platform deployment.
LFM2.5-8B-A1B-MLX	MLX format for Apple Silicon. Optimized for fast inference on Mac devices.
LFM2.5-8B-A1B-DSpark	Speculative decoding drafter (328M). Pair it with this model for ~2.5x faster decoding with identical outputs.

We recommend using LFM2.5-8B-A1B for agentic workflows, tool use, structured outputs, multilingual assistants, and on-device personal-assistant applications. It is not the best fit for heavy programming or knowledge-intensive question answering without retrieval.

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

Because LFM2.5-8B-A1B is a reasoning model, assistant turns contain an explicit chain of thought before the final answer. You can use tokenizer.apply_chat_template() to format your messages automatically.

Tool Use

LFM2.5 supports function calling in four steps:

1. Function definition: Provide the list of tools as a JSON object in the system prompt, or use tokenizer.apply_chat_template() 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: Execute the call and return the result with the tool role.
4. Final answer: LFM2.5 interprets the tool output and returns a plain-text answer addressing the original prompt.

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

LFM2.5-8B-A1B is supported by many inference frameworks. See the Inference documentation for the full list.

Name	Description	Docs	Notebook
Transformers	Simple inference with direct access to model internals.	Link
vLLM	High-throughput production deployments with GPU.	Link
SGLang	High-throughput production deployments with GPU.	Link	—
llama.cpp	Cross-platform inference with CPU offloading.	Link
MLX	Apple's machine learning framework optimized for Apple Silicon.	Link	—
LM Studio	Desktop application for running LLMs locally.	Link	—

[!TIP] ⚡ Faster decoding: attach LFM2.5-8B-A1B-DSpark, a 328M speculative-decoding drafter, for ~2.5x faster decoding in SGLang with exactly the same outputs.

Quick start with Transformers (compatible with transformers>=5.0.0):

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "LiquidAI/LFM2.5-8B-A1B"
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,
)["input_ids"].to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.2,
    top_k=80,
    repetition_penalty=1.05,
    max_new_tokens=8192,
    streamer=streamer,
)

🔧 Fine-Tuning

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

Name	Description	Docs	Notebook
CPT (Unsloth)	Continued Pre-Training using Unsloth for text completion.	Link
CPT (Unsloth)	Continued Pre-Training using Unsloth for translation.	Link
SFT (Unsloth)	Supervised Fine-Tuning with LoRA using Unsloth.	Link
SFT (TRL)	Supervised Fine-Tuning with LoRA using TRL.	Link
DPO (TRL)	Direct Preference Optimization with LoRA using TRL.	Link
GRPO (Unsloth)	GRPO with LoRA using Unsloth.	Link
GRPO (TRL)	GRPO with LoRA using TRL.	Link

📊 Performance

Improvements over LFM2-8B-A1B

Thanks to reasoning, scaled-up pre-training, and large-scale RL, LFM2.5-8B-A1B improves over its predecessor across the board:

Benchmark	LFM2-8B-A1B	LFM2.5-8B-A1B	Δ
AA-Omniscience Index	-78.42	-24.70	+53.62
AA-Omniscience Accuracy	7.33	8.67	+1.34
AA-Omniscience Non-Hallucination Rate	7.46	63.47	+56.01
IFEval	79.44	91.84	+12.40
IFBench	26.00	56.47	+30.47
Multi-IF	58.54	79.93	+21.39
MATH500	74.80	88.76	+13.96
AIME25	20.00	42.53	+22.53
BFCLv3	45.07	64.36	+19.29
BFCLv4	25.52	48.50	+22.98
Tau² Telecom	13.60	88.07	+74.47
Tau² Retail	7.02	39.82	+32.80

Knowledge and instruction following

Model	Parameters	AA-Omni. Index	AA-Omni. Accuracy	AA-Omni. Non-Halluc.	IFEval	IFBench	Multi-IF
LFM2.5-8B-A1B	8B/A1B	-24.70	8.67	63.47	91.84	56.47	79.93
Granite-4.0-H-Tiny	7B/A1B	-75.50	9.37	6.38	82.23	21.28	59.00
Qwen3.5-4B	4B	-51.53	17.20	16.99	87.80	50.38	67.43
Qwen3-30B-A3B-Thinking-2507	30.5B/3.3B	-51.31	18.80	13.87	90.82	51.11	79.04
Gemma-4-E2B-IT	5.1B	-72	7.00	15.05	82.93	33.53	69.70
Gemma-4-E4B-IT	8B	-50.67	8.10	36.06	87.74	39.48	77.58
Gemma-4-26B-A4B-IT	26B/4B	-62.07	14.37	10.75	91.40	47.25	82.06
gpt-oss-20b	21B/3.6B	-49.17	14.57	24.50	86.73	58.65	76.64

Math and agentic workflows

Model	Parameters	MATH500	AIME25	AIME26	BFCLv3	BFCLv4	Tau² Telecom	Tau² Retail
LFM2.5-8B-A1B	8B/A1B	88.76	42.53	50.00	64.79	49.73	88.07	39.82
Granite-4.0-H-Tiny	7B/A1B	59.20	4.93	3.33	56.89	28.52	16.67	18.42
Qwen3.5-4B	4B	80.76	54.28	58.33	71.06	54.01	87.72	71.93
Qwen3-30B-A3B-Thinking-2507	30.5B/3.3B	86.48	71.67	66.67	73.39	50.53	21.93	56.14
Gemma-4-E2B-IT	5.1B	64.00	26	30	56.44	31.91	22.37	18.95
Gemma-4-E4B-IT	8B	65.00	34.33	40.67	57.31	33.92	26.75	42.11

CPU Inference

GPU Inference

LFM2.5-8B-A1B is the fastest model in its size class, reaching 18.5K output tokens per second at high concurrency, over 1.6B tokens per day on a single H100.

📬 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{liquidAI20268BA1B,
  author  = {Liquid AI},
  title   = {LFM2.5-8B-A1B: Personal Assistant On Your Laptop},
  journal = {Liquid AI Blog},
  year    = {2026},
  note    = {www.liquid.ai/blog/lfm2-5-8b-a1b},
}

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

@article{smith2026inplace,
  title   = {In-Place Tokenizer Expansion for Pre-trained LLMs},
  author  = {Smith, Jimmy T.H. and Dakhran, Tarek and Cabrera, Alberto and Lee, Simon S. and Pak, Paul and Tadimeti, Aditya and Seyde, Tim and Labonne, Maxime and Amini, Alexander and Lechner, Mathias},
  journal = {arXiv preprint arXiv:2607.15232},
  year    = {2026},
}