Liquid AI Releases LFM2.5-24B-A2B, a 24B-Parameter MoE for On-Device Agents

Liquid AI said it trained the model inside popular agent harnesses, targeting multi-step tool use rather than open-ended chat. Base and post-trained weights are available on Hugging Face, and the model is listed under the LFM Open License v1.0. It supports 16 languages, including English, Arabic, Chinese, Hindi and Japanese.

The architecture keeps the LFM2 40-layer hybrid design: 30 double-gated short convolution blocks and 10 grouped query attention blocks. Each MoE block contains 64 experts with top-4 routing, according to the model card. Only two things changed from the predecessor LFM2-24B-A2B: context grew from 32,768 tokens to 131,072 tokens, and vocabulary doubled from 65,000 to 128,000 through an in-place tokenizer expansion intended to improve non-Latin tokenization.

Pre-training scaled from 17 trillion to 33 trillion tokens, Liquid AI said. Mid-training ran 2 trillion tokens at 32K context and then 300 billion tokens at 131K context, with tool calls and multi-turn trajectories appearing in both phases. Post-training had five stages: supervised fine-tuning, teacher specialization, non-agentic multi-domain on-policy distillation, agentic multi-domain on-policy distillation, and agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode.

Liquid AI also disclosed reward hacking during teacher specialization. The model searched git history, GitHub, PyPI and Hugging Face for answers and even detected that it was inside a sandbox. The company said it tightened sandboxes and that reported results come from sealed environments.

Liquid AI tested LFM2.5-24B-A2B on 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, a 30B-A3B model. It leads or matches larger MoEs on 5 of 8 agentic and tool-use benchmarks, according to the company. It leads AutomationBench at 27.69 versus 8.72 for Qwen3.6-35B-A3B, ToolSandbox, LiveClaw at 48.09 versus 38.19, and τ³-Bench Banking. BFCLv4 is effectively a tie at 65.01 versus 65.00.

It also tops three instruction-following tests: IFEval at 96.75, IFBench at 74.69 and Multi-IF at 84.38. Larger models win on knowledge and code. Qwen3.6-35B-A3B leads GPQA Diamond at 83.69 versus 74.42, SWE-Bench Verified at 66.20 versus 51.00, WebArena Verified at 37.62 versus 26.75 and PinchBench at 89.93 versus 83.18. Liquid AI's bottom-line assessment is that the model is fast and cheap for local tool calling and instruction following, but weaker on knowledge and coding and should be paired with retrieval.

On devices, LFM2.5-24B-A2B decodes 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, Liquid AI said. The measurements used llama.cpp, Q4_K_M quantization and a 4K-token input. Liquid AI claims up to 1.7 times the speed of the next-best compared MoE, with the lowest memory use, but exact RAM requirements were not disclosed. For serving, it passes 21,000 output tokens per second on a single NVIDIA B300, using SGLang 0.5.22, BF16, 1,024 input tokens and up to 256 output tokens.

Day-one support covers Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio and ONNX, with FP8, GGUF, ONNX and MLX 4/5/6/8-bit builds. MLX requires mlx-lm 0.32.0 or later because earlier versions route experts incorrectly, according to the report. Recommended sampling is temperature 0.1, top_k 50 and repetition penalty 1.05. A DSpark speculative-decoding draft model is listed as coming soon.