Liquid AI has released the open model LFM2.5-24B-A2B with 24 billion parameters for local operation on laptops and mini-PCs. The release took place on October 13, 2026, and the model is available on Hugging Face. This is the largest and latest model of the LFM2.5 family.
The model is of the MoE type: a total of 24 billion parameters, but about 2 billion, specifically 2.45 billion, are activated per token. Context is 131,072 tokens, vocabulary is 128,000 entries, supporting 16 languages, including English, Arabic, Chinese, Hindi, and Japanese. The model is focused on multi-step tool usage rather than open chat: each response starts with a think block.
The architecture is a 40-layer hybrid: 30 short convolution blocks with double gating and 10 group attention blocks. Each MoE block contains 64 experts with top-4 routing. Compared to its predecessor LFM2-24B-A2B, the model differs in context growth from 32,768 to 131,072 tokens and vocabulary from 65,000 to 128,000.
Pretraining increased from 17 to 33 trillion tokens. Then mid-training: 2 trillion tokens with 32K context and another 300 billion with 131K context. Post-training includes five stages, including reinforcement agent training. The company revealed a case of reward hacking: the model searched for answers in git, on GitHub, PyPI, and Hugging Face and determined what was in the sandbox. Sandboxes were tightened.
According to Liquid AI, the model leads or ties for first place in 5 out of 8 agent tests against larger models, including Qwen3.6-35B-A3B. In AutomationBench – 27.69 versus 8.72. In LiveClaw – 48.09 versus 38.19. In instruction-following tests: IFEval 96.75, IFBench 74.69, Multi-IF 84.38. Larger models win in knowledge and code: GPQA Diamond 83.69 versus 74.42, SWE-Bench Verified 66.20 versus 51.00.
On Apple M5 Max, the model outputs 184 tokens per second, on AMD Ryzen AI Max+ 395 – 114. Claimed to be up to 1.7 times faster than comparable MoE with lower memory consumption. For servers – over 21,000 output tokens per second on a single NVIDIA B300. Support from day one: Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio, and ONNX. For MLX, version mlx-lm 0.32.0 or newer is required.
The API costs $0.10 per 1 million input tokens and $1.00 per 1 million outputs. Weights are under the LFM Open License v1.0 based on Apache 2.0: free commercial use ceases at an annual revenue of $10 million.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.
Related articles
