Liquid AI releases open-weight LFM2.5-24B-A2B for on-device AI agents

Liquid AI on October 13, 2026 released LFM2.5-24B-A2B, a Mixture-of-Experts reasoning model with 24 billion total parameters but only 2 billion active for each token. The company built it to power fast, private agents that run on users' own hardware. Both the model and a base version, LFM2.5-24B-A2B-Base, are available as open weights on Hugging Face.

According to Liquid AI, the model always writes out an explicit chain of thought before answering, can call tools, and handles multi-step tasks inside popular agent harnesses. Because so few parameters fire per token, each reasoning step is inexpensive, which the company says makes it practical to keep agents running continuously on local machines.

Architecture and training

The release keeps the architecture of its predecessor, LFM2-24B-A2B: 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. The context window expands from 32K to 131K tokens, and the vocabulary doubles from 65K to 128K tokens to improve handling of non-Latin scripts. Pre-training grew from roughly 17 trillion to about 33 trillion tokens, followed by a five-stage post-training pipeline that ends with reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.

Benchmarks and speed

In Liquid AI's own tests across 16 benchmarks, the model scored 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters. The company says it more than tripled the next-best score on AutomationBench and led all three instruction-following benchmarks. Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning were also in the comparison.

The results are not uniformly favourable. Larger rivals stay ahead on knowledge-heavy tests such as GPQA Diamond and HLE and on coding tasks: on SWE-Bench Verified the model scored 51.00 versus 66.20 for Qwen3.6-35B-A3B. Liquid AI recommends pairing it with retrieval or web search for knowledge-intensive work. All figures are self-reported and have not been independently verified.

On performance, the company reports 184 tokens per second on an Apple M5 Max, 114 on an AMD Ryzen AI Max+ 395, and over 21,000 output tokens per second at high concurrency on a single NVIDIA B300 GPU. Support for llama.cpp, MLX, vLLM, SGLang and ONNX is available from launch. The release reflects a broader push toward capable agents that run locally rather than relying on cloud services.