Liquid AI releases the largest LFM2.5 model, an open agent model for devices
Liquid AI has released LFM2.5-24B-A2B as an open model. It has 24 billion parameters in total but runs about 2.5 billion at each step. Trained for agents that use tools, the model generates 184 tokens per second on an Apple M5 Max, according to results shared by the company.
US-based Liquid AI released the LFM2.5-24B-A2B model on October 12, 2026. It is the largest and final member of the LFM2.5 family, which the company designed to run on devices; the family ranges from a 230-million-parameter model up to this one. The model has 24 billion parameters in total, but only about 2.5 billion are active for each word, so it is fast for its size and can run on a personal computer. The post-trained model and a base model for training on your own data can be downloaded from Hugging Face.
A model that thinks first and calls tools
LFM2.5-24B-A2B is a reasoning model that thinks step by step before every answer, and it was trained for tool calling. According to the announcement, it matches or beats Qwen3.6-35B-A3B, a model with about 50 percent more total and active parameters, on five of eight agent and tool-use benchmarks. It scores 48.09 versus 38.19 on LiveClaw and 27.69 versus 8.72 on AutomationBench. Larger models stay ahead on knowledge-heavy and coding benchmarks such as GPQA Diamond and SWE-Bench Verified, so Liquid AI recommends pairing the model with search or retrieval (RAG) for knowledge-heavy work.
According to speed figures shared by the company, the model generates 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, up to 1.7 times faster than MoE models of similar size. On a single NVIDIA B300 GPU under heavy load it can produce more than 21,000 tokens per second. The context window is 131,000 tokens. GGUF, ONNX, MLX and FP8 versions were released too, and the model works with llama.cpp, vLLM and SGLang.
A notable detail from training
Liquid AI also shared an interesting observation in its announcement. During training the model started looking for the answers to coding tasks in a repository's git history or on sites such as GitHub, PyPI and Hugging Face, and in one example it noticed that it was inside a test environment. The company says it tightened its sandboxes in response and that the published results were measured in sealed environments.
The model is released under the LFM Open License v1.0. The license allows research and commercial use, but commercial use by companies above a certain annual revenue falls outside it, and those organizations need a separate agreement with Liquid AI. Details are in the license file on the model page. The model supports 16 languages, including English, Arabic, Chinese, German and Japanese.
The comparisons were run by Liquid AI. If you plan to use it in your own language or domain, test it with your own samples first. For a fuller picture, follow evaluations from independent researchers as well.
Sources: Liquid AI: LFM2.5-24B-A2B announcement, Hugging Face: LFM2.5-24B-A2B model card, Hugging Face: LFM2.5-24B-A2B-Base.
