Liquid AI’s open-weight LFM2.5-24B-A2B model has 2 billion active parameters and is designed for on-device reasoning and tool use.
Liquid AI has released LFM2.5-24B-A2B, an open-weight model designed to power reasoning and tool-using agents on local hardware. The model has 24 billion total parameters, with 2 billion active per token, and is available alongside a base version on Hugging Face. Liquid AI says it can be used with agent frameworks and common inference software, including llama.cpp, MLX, vLLM, SGLang, and ONNX.
The model keeps the 40-layer mixture-of-experts architecture of its predecessor, LFM2-24B-A2B, but extends the context window from 32,000 to 131,000 tokens. Liquid AI also expanded the tokenizer vocabulary from 65,000 to 128,000 tokens, with the stated aim of improving non-Latin text handling. The company describes this release as reasoning-focused: its training process includes tool-use examples and multi-turn agent trajectories, followed by stages of supervised fine-tuning, distillation, and agentic reinforcement learning.
In Liquid AI’s evaluations across 16 benchmarks, LFM2.5-24B-A2B scored 48.09 on LiveClaw, compared with 38.19 for Qwen3.6-35B-A3B. It also led the compared models on AutomationBench, scoring 27.69, and matched or narrowly exceeded Qwen3.6-35B-A3B on BFCLv4, with scores of 65.01 and 65.00 respectively. The company’s results show weaker scores than some larger competitors on several knowledge and coding tests, including GPQA Diamond and SWE-Bench Verified; Liquid AI suggests pairing the model with retrieval or web search for knowledge-intensive use.
For speed, Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395. On a single NVIDIA B300, it reports more than 21,000 output tokens per second at high concurrency. These are company-reported results; the article says the compared models were evaluated using vLLM and each benchmark’s recommended generation settings.