Liquid AI's LFM2.5-24B Is a 24 Billion Parameter Model That Only Thinks With 2 Billion at a Time

Liquid AI LFM2.5 puts a big model on local hardware

Liquid AI released LFM2.5-24B-A2B on October 13, and it is built for a specific job: running AI agents on your own machine. The model has 24 billion total parameters, but it only activates about 2.45 billion of them per token. That mixture-of-experts trick is the whole point. You get a large model's capacity without paying a large model's compute bill on every single token.

The release closes out the LFM2.5 family, which ranges from tiny 230 million parameter models up to this one. Both base and post-trained weights are on Hugging Face, so developers can download, customize, and deploy it themselves. It is text only and reasoning only, which means every answer starts with an explicit chain of thought before the final response. That slows things down a little, but it is exactly what multi-step agent work needs.

What changed from the last version

The architecture stays familiar: 40 layers, 64 experts per mixture-of-experts block, top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. What changed is scale. The context window jumped from 32,768 tokens to 131,072 tokens, and the vocabulary doubled from 65,000 to 128,000 tokens, which Liquid AI says improves tokenization for non-Latin scripts. The model supports 16 languages, including English, Arabic, Chinese, Hindi, and Japanese.

Training scaled up too. Pre-training grew from roughly 17 trillion tokens to about 33 trillion, followed by mid-training focused on reasoning, math, code, instruction following, and multilingual text. Post-training ran through five stages, ending with reinforcement learning inside real agent harnesses: OpenClaw, Hermes Agent, Pi, and OpenCode. That last detail matters more than the parameter count. This model was not just trained to answer questions. It was trained inside the tools it is supposed to drive.

Liquid AI LFM2.5 benchmarks: strong on agents, weaker on knowledge

Liquid AI's own benchmarks tell a split story, and it is worth reading both halves. On agentic tests, the model shines. AutomationBench scored 27.69, more than three times Qwen3.6-35B-A3B at 8.72. LiveClaw came in at 48.09 against 38.19 for the same rival. Instruction following looked strong too, with IFEval at 96.75.

On knowledge and coding benchmarks, larger rivals keep the lead. GPQA Diamond scored 74.42 against 83.69 for Qwen3.6-35B-A3B. SWE-Bench Verified with OpenCode scored 51.00 against 66.20. WebArena Verified scored 26.75 against 37.62. Liquid AI itself advises pairing the model with retrieval or web search for knowledge-heavy work. That is honest advice, and it tells you how to use this model: let it drive the workflow, and give it tools to look things up.

Who this is actually for

Three groups should pay attention:

• Developers building local agents. If your agent loops through tool calls all day, a model that runs locally with a small active footprint cuts latency and keeps data on the machine.
• Teams with privacy constraints. A downloadable model means prompts, files, and tool results never have to leave your hardware. For internal tools, that can matter more than raw benchmark scores.
• Tinkerers with decent GPUs. Independent analysis suggests a quantized build with 128K context can fit on a 16 GB card. That is not a casual laptop spec, but it is within reach of a serious workstation.

The comparison with Moonshot's Kimi K3.1 Pro is instructive. Kimi is the stronger model overall, but it is a giant that most people will only ever touch through an API. Liquid AI's model is the one you can actually run yourself. That trade is the story of open weights in late 2026: the frontier stays in the cloud, while the usable models keep moving onto your desk.

The honest caveats

A few things to keep in mind. First, a 2 billion active parameter footprint does not mean 2 billion parameter memory usage. The full 24 billion parameters still have to live somewhere, and quantized builds are a compromise, not magic. Second, reasoning-only output means more tokens per answer, so the per-token savings get partly eaten by longer responses. Third, these are vendor benchmarks, run by Liquid AI on its own terms. Independent testing will sharpen the picture.

None of that kills the appeal. LFM2.5-24B-A2B is a serious, downloadable agent model with a clear purpose, and it arrives in a month crowded with open releases, from DeepSeek's V4.1-Flash to Mistral's trillion-parameter model. If you want an agent that lives on your machine, this is one of the best options you can download today.