Run fast, private agents on-device with LFM2.5-24B-A2B

Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model aimed at running agents locally, announcing it in a Hugging Face post dated October 13, 2026. The model has 24 billion total parameters but activates only 2 billion per token, which the company says makes each reasoning token cheap enough to run agents continuously on personal hardware. A base version is also available.

The model keeps its predecessor's architecture of 40 layers and 64 experts per mixture-of-experts block, but now always produces a chain of thought before answering. Its context window grows from 32K to 131K tokens, its vocabulary doubles to 128K tokens, and pre-training rose from about 17 trillion to 33 trillion tokens. Post-training included reinforcement learning inside agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.

In Liquid AI's own tests across 16 benchmarks, the model led or matched Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning on five of eight tool-use and agentic measures, scoring 48.09 on LiveClaw against Qwen's 38.19, and on all three instruction-following tests. Its results were weaker on knowledge-heavy and coding benchmarks: it scored 51.00 on SWE-Bench Verified versus Qwen's 66.20, and the company suggests pairing it with retrieval for knowledge work.

Liquid AI reports speeds of 184 tokens per second on an Apple M5 Max, 114 on a Ryzen AI Max+ 395, and over 21,000 at high concurrency on a single NVIDIA B300 GPU, with support for llama.cpp, MLX, vLLM, SGLang and ONNX.