LFM2.5-2.6B

LFM2.5-2.6B is Liquid AI's August 4, 2026 open-weight release: a 2.6-billion-parameter model trained for agentic work — planning a task, calling tools, reading the results, and taking the next step — at a size that fits on a phone, a laptop, or an edge box rather than a GPU server. It sits at the top of the LFM2 family, which spans sub-billion checkpoints up through this one, and it is distributed as weights you download and run yourself. There is no hosted endpoint and no per-token bill.

The audience is developers building local agents: offline assistants, on-device automation, privacy-constrained pipelines in health or finance where shipping user data to a hosted API is the thing that kills the project, and anyone who wants an agent loop that keeps working in airplane mode. What separates it from the usual open small model is what it was tuned for. Most 2–3B open checkpoints are optimised for chat and completion, and function calling is bolted on afterwards with mixed reliability; Liquid is positioning tool calling and multi-step execution as the primary job. The LFM2 line also uses a hybrid backbone — short convolution blocks interleaved with grouped-query attention rather than a uniform transformer stack — chosen for throughput on CPUs and mobile NPUs, not for leaderboard scores on a datacentre GPU. Liquid pairs the weights with its own edge stack (the LEAP SDK for iOS and Android, plus the Apollo app for trying models locally), which is more deployment scaffolding than most labs ship with a Hugging Face repo.

The honest caveats: 2.6B is 2.6B. This will not hold a long-horizon coding session the way a frontier hosted model does, and if your agent needs deep multi-file reasoning you are picking the wrong weight class. The performance claims are vendor-published and, this soon after release, largely un-replicated by third parties — benchmark it against your own task before committing. And check the licence text yourself: Liquid has historically released LFM weights under its own LFM Open License rather than plain Apache 2.0, with terms that turn on company revenue. That distinction matters a lot if you plan to embed this in something you sell.