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

LFM2.5-24B-A2B is a compact 24-billion-parameter reasoning model designed for fast, private agents running on-device. Unlike its predecessor, it focuses solely on reasoning, providing an explicit chain of thought before delivering its final answer. The model boasts a significant increase in context window size from 32K to 131K tokens and expands its vocabulary from 65K to 128K tokens.

LFM2.5-24B-A2B builds upon the LFM2-24B-A2B architecture, maintaining 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and GQA.

Pre-training was expanded from approximately 17 trillion to 33 trillion tokens, with additional mid-training phases incorporating 2 trillion tokens geared towards reasoning, math, code, instruction following, and multilingual text. Following that, 300 billion tokens were added at the 131K context size, emphasizing long-context data. The model undergoes five stages of post-training to become a fully functional agent.

In benchmark evaluations, LFM2.5-24B-A2B was compared to Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning across 16 benchmarks covering tool use and agentic workflows, knowledge, instruction following, math, and coding. Despite having the fewest parameters, it performed competitively with larger models that possess up to 50% more total and active parameters.

It significantly outperformed the competition on five out of the eight tool-use and agentic benchmarks, including AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking. Additionally, it excelled on three instruction-following benchmarks.

While larger models maintain an advantage on knowledge-heavy benchmarks like GPQA Diamond, HLE, and AA-Omniscience, as well as coding benchmarks such as SWE-Bench Verified (OpenCode) and SciCode, LFM2.5-24B-A2B is well-suited for agents that require reasoning, tool calls, and reliable on-device operation. It is compatible with various inference ecosystems, including llama.cpp, MLX, vLLM, SGLang, and ONNX.

LFM2.5-24B-A2B provides the fastest decoding speed among its peers, achieving 184 tokens per second on an M5 Max and 114 tokens per second on a Ryzen AI Max+. It is available as an open-weight model on Hugging Face and can be served locally through popular agent harnesses such as Hermes Agent, llama.cpp, MLX, vLLM, SGLang, and LM Studio.