Liquid AI releases open-weight model built for on-device agents

Liquid AI has released LFM2.5-24B-A2B, an open-weight model aimed at running AI agents on devices and other constrained hardware. It pairs a 131,072-token context window with day-one support across several inference tools, while the company’s own comparisons show stronger results on some agent tasks than on knowledge-heavy and coding benchmarks.

Watch Desk analysis

What happened

The model has 24 billion total parameters, with 2.45 billion active, and is the largest model in Liquid AI’s LFM2.5 family. Its context window is four times the 32,768 tokens of its predecessor, and its vocabulary has doubled to 128,000 tokens. The release is available in native, FP8, GGUF, ONNX and MLX-4bit formats, with support for llama.cpp, MLX, vLLM, SGLang and ONNX.

Liquid AI reports that the model led or matched its comparison group on five of eight tool-use and agent benchmarks, and led on all three instruction-following benchmarks. Its tables also show larger comparison models ahead on some knowledge and coding tests. The company recommends adding retrieval or web search for knowledge-heavy work. Unite.AI reports that hosted API pricing is $0.10 per million input tokens and $1 per million output tokens.

Why it matters

This is a substantial open-weight release for developers who want to run agent workflows on local or edge hardware, rather than sending every task to a cloud service. The varied formats and inference support lower the friction of trying it on different setups; the benchmark split offers a more useful guide than a single headline score.

The performance and speed figures are Liquid AI’s own measurements, not independent comparisons. That matters when deciding whether the model fits a real workload, but it need not stop developers from giving it a practical test.

Our read

The appealing bit is the combination of open weights, broad runtime support and a model designed with agent use in mind. The sensible next step is to try it on the tasks you actually care about, especially if those tasks rely on factual knowledge or substantial coding. A model can be quick on the local machine and still need a capable retrieval system beside it.

What to watch

• Whether independent tests reproduce the company’s speed and benchmark results.
• How well the model performs in real agent workflows on edge hardware.
• Whether Liquid AI releases the promised speculative-decoding draft model for SGLang.

Discussion spark: Would you run an open-weight model locally for agent tasks if it did well at tool use but needed retrieval support for knowledge-heavy work, or would you keep those jobs in the cloud?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.