Liquid AI unveils LFM2.5-24B-A2B: a high-performance moe model for local agents

Liquid AI has launched LFM2.5-24B-A2B, a 24B-parameter Mixture-of-Experts (MoE) model designed for on-device agents running on laptops and PCs. The model activates approximately 2B parameters per token and is the final member of the LFM2.5 series. It is optimized for multi-step tool use rather than open-ended chat, with base and post-trained weights available on Hugging Face. The model supports 16 languages, including English, Arabic, Chinese, Hindi, and Japanese, and features a 40-layer hybrid architecture with 30 double-gated short convolution blocks and 10 grouped query attention blocks. LFM2.5-24B-A2B improves upon its predecessor, LFM2-24B-A2B, with an expanded context window from 32,768 to 131,072 tokens and a doubled vocabulary from 65K to 128K. Pre-training data increased from 17T to 33T tokens, with mid-training phases involving 2T and 300B tokens at different context lengths. Liquid AI also noted the model's ability to detect it was running in a sandbox environment, highlighting enhanced security measures. The model outperformed competitors in several benchmarks, including AutomationBench, ToolSandbox, LiveClaw, and τ³-Bench Banking. It also excelled in instruction-following tests, achieving scores of 96.75 on IFEval, 74.69 on IFBench, and 84.38 on Multi-IF. On-device performance reached 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+.