Liquid AI Launches LFM2.5-24B-A2B On-Device Model

Liquid AI has launched LFM2.5-24B-A2B, an open-weight reasoning model featuring 24 billion total parameters with only 2 billion active parameters per token. The mixture-of-experts model uses 40 layers, 64 experts per block with top-4 routing, and a hybrid architecture of gated short convolutions and grouped-query attention. Built for continuous on-device operation, it expands its context window from 32K to 131K tokens and doubles its vocabulary size from 65K to 128K tokens. Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are available on Hugging Face with day-one ecosystem support for llama.cpp, MLX, vLLM, SGLang, ONNX, and transformers>=5.0.0.

Pre-training was scaled from ~17T to ~33T tokens, followed by mid-training with 2T tokens at 32K context and 300B tokens at 131K context. Post-training used multi-domain on-policy distillation across non-agentic and agentic stages, alongside agentic reinforcement learning within harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode. Evaluated across 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning using vLLM, LFM2.5-24B-A2B scored 48.09 on LiveClaw versus Qwen3.6-35B-A3B's 38.19. It led or matched competitors on five of eight tool-use benchmarks—including AutomationBench, where it scored over three times higher than the next-best model, as well as BFCLv4, ToolSandbox, and τ³-Bench Banking—and led on all three instruction-following benchmarks.

Inference tests show decoding speeds of 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, delivering up to 1.7 times faster performance than competitors on edge hardware. On a single NVIDIA B300 GPU, it reaches over 21K output tokens per second at high concurrency. While larger models retain an edge on knowledge-heavy benchmarks like GPQA Diamond, HLE, and AA-Omniscience, as well as coding benchmarks like SWE-Bench Verified (OpenCode) and SciCode, developers can pair LFM2.5-24B-A2B with external retrieval or web search to build fast, low-cost local agents.