Most agent models ask you to choose between capability and cost. Liquid AI's LFM2.5-24B-A2B tries to remove that trade-off: it is a large open-weight model that only wakes up a small slice of itself for each token.
The idea in plain terms
The headline is simple. The model stores 24 billion parameters, but activates only about 2 billion per token. That is the Mixture-of-Experts trick: keep a library of specialised sub-networks, route each token to a few of them, and get much of the knowledge of a big model at roughly the running cost of a small one.
It is also built as a reasoning model. It thinks before it acts, which matters for long agent work where a plan, a tool call, a result check and a revised plan all have to stay coherent.
What changed from the previous version
Three upgrades stand out for builders:
- A context window of 131,072 tokens, up from 32K, so long sessions, documents and agent runs fit in one place.
- A vocabulary doubled to 128K tokens, improving multilingual handling across 16 languages.
- A much larger training budget, with post-training tuned for real agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.
Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX, which makes it practical to try locally rather than only through an API.
Where it shines, and where it does not
On Liquid AI's own benchmarks, the model leads on tool use and instruction following. AutomationBench and LiveClaw are the eye-catching results, and IFEval and IFBench point the same way.
Treat those numbers as vendor claims, not independent proof. On knowledge-heavy and coding tests, larger rivals still lead. That split is useful: this is a model for doing things, not necessarily for answering exam-style questions from memory.
Should you run it?
If you are building a local agent, working with private data, or paying a recurring bill for an always-on assistant, it is worth testing. A capable model that runs on your own hardware changes the economics of agents that watch, check and act all day.
If you only need the smartest single answer to a question, a frontier cloud model is still the simpler choice. And on a modest laptop, check memory first: the full model is happiest on high-end machines, even if the per-token cost is low.
For the full breakdown, including the benchmark table, hardware notes and the verdict, read the full guide on NeurAITool.
Top comments (0)