Liquid AI’s new model promises faster private agents on your device

Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model designed to run AI agents locally on consumer and enterprise hardware. Liquid AI announces in an official blog post that the model contains 24 billion parameters but activates only 2 billion for each token it processes.

That mixture-of-experts design aims to reduce the computing work required during inference. According to the company, the model reaches 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395. Liquid AI also reports more than 21,000 output tokens per second on a single Nvidia B300 GPU under high concurrency.

The model is the largest release in Liquid AI’s LFM2.5 family. It has a context window of 131,072 tokens, up from 32,768 in the previous version, and a vocabulary expanded from 65,000 to 128,000 tokens. The company says the larger vocabulary improves processing of non-Latin languages.

Focus on tools and agent workflows

LFM2.5-24B-A2B is a reasoning-only model, meaning it produces internal step-by-step reasoning before delivering an answer. Liquid AI says it trained the model on tool calls, multi-turn interactions, coding, math, instruction following, and long-context tasks. Its final training stage uses agent environments including OpenClaw, Hermes Agent, Pi, and OpenCode.

In company-provided benchmark results, the model performs strongly on several tool-use tests. It scores 27.69 on AutomationBench, compared with 8.72 for Qwen3.6-35B-A3B, and 48.09 on LiveClaw, compared with 38.19. It also leads the listed models on three instruction-following benchmarks.

However, Liquid AI reports that larger models retain an advantage in knowledge-heavy evaluations such as GPQA Diamond and HLE, as well as software engineering tests including SWE-Bench Verified. The company recommends retrieval or web search for knowledge-intensive work.

Base and post-trained versions are available through Hugging Face and Liquid AI’s Playground. The release supports llama.cpp, MLX, vLLM, SGLang, and ONNX, enabling local deployment across Apple, AMD, Qualcomm, and Nvidia hardware.