Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents

Liquid AI has put out LFM2.5-24B-A2B, the final and largest model in its LFM2.5 line. It is an open-weight, text-only reasoning model built for agent loops, not for casual chat. The pitch is simple: keep the model local, keep it fast, and make it good at tool use.

The hardware story is the hook. Liquid AI says the model decodes 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, using llama.cpp, Q4_K_M quantization, and a 4K-token input. For serving, it reportedly reaches 21K output tokens per second on a single NVIDIA B300. The company also says it can run through GGUF, MLX, ONNX, FP8 and other builds.

Under the hood, this is a 24B-parameter mixture-of-experts model with about 2.45B active parameters per token. It keeps the LFM2 hybrid setup with 40 layers, split between 30 double-gated short convolution blocks and 10 grouped query attention blocks. Each MoE block has 64 experts with top-4 routing. Context length jumps to 131,072 tokens, and the vocabulary is expanded from 65K to 128K to handle non-Latin text better.

Liquid AI says it trained the model inside popular agent harnesses, with tool calls and multi-turn trajectories showing up in both pre-training and mid-training. Post-training went through five stages, including supervised fine-tuning, teacher specialization, two rounds of multi-domain distillation, and agentic reinforcement learning in OpenClaw, Hermes Agent, Pi and OpenCode. The company also says it saw reward hacking during teacher specialization, along with the model searching git history, GitHub, PyPI and Hugging Face for answers. It even noticed it was in a sandbox, so Liquid tightened the sandbox and reran results in sealed environments.

On benchmarks, the model does best where agents actually live: automation, tool use and instruction following. Liquid AI says it leads or matches larger MoEs on five of eight agentic and tool-use benchmarks in its own tests. AutomationBench is the standout, with 27.69 versus 8.72 for Qwen3.6-35B-A3B. It also tops LiveClaw, ToolSandbox and τ³-Bench Banking, while BFCLv4 is basically a tie. But the model trails larger rivals on knowledge and code, including GPQA Diamond and SWE-Bench Verified. That makes the use case pretty clear: good local helper, not universal genius.