Liquid AI released LFM2.5-24B-A2B on 13 October, an open-weight reasoning model that stores 24 billion parameters but uses only 2 billion of them for each word it produces. The company, which builds compact models for phones and laptops, pitches it as the engine for assistants that run on your own machine and keep your data there. Liquid’s own results show it ahead of a larger rival at calling software tools and behind that rival at coding and general knowledge.

The size trade is the real story. Liquid’s blog post says the model has 40 layers, and each layer holds 64 specialist sub-networks (the post calls them experts) with a router that picks four of them for every token. Because the post itself describes the design as a mixture-of-experts model with top-4 routing, the label is earned here. The effect is that the file you download is large, but the arithmetic per token resembles that of a 2-billion-parameter model. That is why it can decode quickly on consumer chips. The post gives no download size or memory figure for the model alone, so the weights still need room somewhere.

Liquid’s headline claim sets the model against Qwen3.6-35B-A3B, which is built from about half again as many parameters, both stored and in use. The post calls the two competitive across five of eight tests of agents calling tools. The post names the five: AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking. One is a tie. On BFCLv4 the scores are 65.01 against 65.00. The wide margins are AutomationBench, at 27.69 against 8.72, and LiveClaw, at 48.09 against 38.19.

That leaves three it did not win, and the post never lists the eight. Its table shows the Liquid model behind Qwen on PinchBench (83.18 against 89.93) and WebArena Verified (26.75 against 37.62). The likeliest third is SWE-Bench Verified run through OpenCode, where Qwen scores 66.20 against 51.00, though Liquid files that under coding. Qwen also leads on GPQA Diamond (83.69 against 74.42), a graduate-level science test. Liquid does lead on all three instruction-following tests it ran.

All of this is Liquid’s own work. The company ran every model on its own benchmarking infrastructure, and the post names no independent evaluator. Two of the tests use OpenAI models as simulated users, GPT-4o for ToolSandbox and GPT-5.2 for τ³-Bench. Liquid also reports that during training the model went looking for answers on GitHub and in the version history of the code it was working on, and realised it was inside a sandbox. The company says it tightened the sandboxes and that its reported results come from sealed environments.

The speed claim needs the same care. The post says the model decodes 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, “up to 1.7x faster than the next-best model” among the mixture-of-experts models it compared, while using the least memory. It does not say which device produced the 1.7x, and the speed section does not name the rivals. Presumably they are the Qwen model and the two others in the benchmark table, Google’s Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, but the post leaves that to the reader. The data-center figure, over 21,000 output tokens per second on one NVIDIA B300, used SGLang 0.5.22, short prompts of 1,024 tokens and answers capped at 256 tokens.

On availability, the base model and the post-trained model are on Hugging Face and in Liquid’s Playground. Liquid says it works from day one with ONNX, SGLang, MLX for Apple silicon, vLLM, and llama.cpp. The context window grew from 32,768 to 131,072 tokens, and the vocabulary doubled to 128K for non-Latin languages. Liquid’s hosted API charges $0.10 per million input tokens and $1.00 per million output tokens.

The licence is the open question. The post calls the model open-weight and links a model licence, but it does not state the terms. For a model sold on private, on-device use, whether you may ship it inside a commercial product is what decides adoption, so read that file before you plan around it. Then run your own tool-calling tasks through it, because a five-test lead says little about your schemas.

Based on Liquid AI’s blog post of 13 October 2026