Liquid AI has released LFM2.5-24B-A2B, a 24B-parameter mixture-of-experts reasoning model that activates only 2B parameters per token. According to a Hugging Face blog post from the Liquid AI team, the model is designed for on-device agents that need to reason, call tools, and work through multi-step tasks. It builds on LFM2-24B-A2B with the same architecture but adds explicit chain-of-thought reasoning, extends context from 32K to 131K tokens, and doubles vocabulary to 128K tokens.
The team reports that the model is competitive with Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters. It leads or matches on five of eight tool-use and agentic benchmarks, including AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking, and tops all three instruction-following benchmarks tested. Larger models still hold an edge on knowledge-heavy and coding benchmarks such as GPQA Diamond, HLE, SWE-Bench Verified, and SciCode.
On hardware, the model reaches 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395, and over 21K output tokens/s at high concurrency on a single NVIDIA B300 GPU. It ships with day-one support in llama.cpp, MLX, vLLM, SGLang, and ONNX. The post also details a five-stage post-training pipeline, including supervised fine-tuning, teacher specialization, two distillation stages, and agentic reinforcement learning inside harnesses such as OpenClaw and OpenCode.