Liquid AI releases LFM2.5-24B-A2B, an open reasoning model with 2B active parameters for on-device agents
Liquid AI published LFM2.5-24B-A2B and its base model as open weights on Hugging Face. The mixture-of-experts model activates about 2 billion of its 24 billion parameters per token, has a 131K-token context window and is aimed at agents that run on local hardware.
Photo: Kenneth C. Zirkel / Wikimedia Commons, CC BY 4.0. Hugging Face and Liquid AI logos via Wikimedia Commons (public domain).
Liquid AI announced LFM2.5-24B-A2B on the Hugging Face blog on October 13. The company describes it as a 24-billion-parameter reasoning model with just 2 billion active parameters per token, built for fast, private agents that run on-device. It always produces a chain of thought before answering, calls tools and works through multi-step tasks inside popular agent harnesses.
The model keeps the architecture of its predecessor, LFM2-24B-A2B: 40 layers, 64 experts per mixture-of-experts block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. Liquid says the context window grows from 32K to 131K tokens, the vocabulary doubles from 65K to 128K tokens, and pre-training was scaled from about 17 trillion to about 33 trillion tokens. Post-training runs in five stages, ending with agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.
In Liquid's own evaluations against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, LFM2.5-24B-A2B scored 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, and led or matched the others on five of eight tool-use and agentic benchmarks. Liquid acknowledges that larger models keep an edge on knowledge-heavy tests such as GPQA Diamond and on coding benchmarks such as SWE-Bench Verified, and suggests pairing the model with retrieval or web search for knowledge-heavy work.
Liquid reports decoding speeds of 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, and more than 21,000 output tokens per second at high concurrency on a single NVIDIA B300 GPU. Both the model and LFM2.5-24B-A2B-Base are available on Hugging Face, with day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX.