Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model meant to run AI agents on laptops and workstations instead of in the cloud. The company described the model in a Hugging Face blog post dated October 13, 2026, and published both the instruction-tuned version and a base checkpoint on its Hugging Face organization.
The design leans on sparsity. The model holds 24 billion parameters in total, yet a mixture-of-experts router activates only about 2 billion of them for each token, which keeps the cost of every generated token low. That matters here because LFM2.5 is a reasoning-only model: it always writes out a chain of thought before answering, so it produces many tokens per task. Liquid AI kept the architecture of its earlier LFM2-24B-A2B, with 40 layers, 64 experts per block, top-4 routing and a hybrid of gated short convolutions and grouped-query attention, but widened the context window from 32,000 to 131,000 tokens and doubled the vocabulary to 128,000 tokens to handle non-Latin scripts better.
Most of the change is in training. Pre-training roughly doubled, from about 17 trillion to about 33 trillion tokens, followed by long-context mid-training. Post-training ran in five stages, ending with reinforcement learning carried out inside real agent harnesses, among them OpenClaw, Hermes Agent, Pi and OpenCode, so the model learns to call tools in the environments where people will actually use it.
Liquid AI's own evaluation compares the model with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning across 16 benchmarks, all served through vLLM. By the company's figures, LFM2.5 leads or ties on five of eight tool-use and agentic tests, scoring 48.09 on LiveClaw against 38.19 for the larger Qwen model and 27.69 on AutomationBench, more than triple the next result, and it tops all three instruction-following benchmarks. The same table shows clear weaknesses: Qwen3.6 scores 66.2 on SWE-Bench Verified to LFM2.5's 51.0 and leads on GPQA Diamond, and the larger rivals stay ahead on knowledge-heavy tests. Liquid AI itself recommends pairing the model with retrieval or web search for such workloads.
Speed is the other selling point. Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395. By its count that is as much as 1.7 times quicker than the next model in the comparison, with the smallest memory footprint, and a single NVIDIA B300 serving many requests at once can push out above 21,000 tokens per second. Day-one runtimes include llama.cpp, MLX, vLLM, SGLang and ONNX.
The release adds a capable option for developers who want agents to keep data on their own machines, a niche where NVIDIA's Nemotron 3.5 Bolt also announced this week is competing. The licence deserves attention, though. LFM2.5 ships under the LFM Open License v1.0, which grants free use but excludes commercial use by organisations with annual revenue of $10 million or more, so larger companies cannot treat it like an Apache-licensed model.
All benchmark and speed figures come from Liquid AI and have not yet been reproduced independently.