What actually shipped

Most model releases this month have been closed, hosted and priced per token. This one is a download. Liquid AI has put both the finished model and its base version on Hugging Face, with day-one support in the local serving tools people already use. The design is the familiar mixture-of-experts trade, pushed toward the edge: the model stores 24 billion parameters but activates 2.45B per token (rounded to 2B in Liquid's marketing), and uses the saving to run continuously on hardware a user owns.

Two changes separate it from the LFM2 version that came before. The context window quadruples, from 32K to 131K tokens, and the model now always reasons out loud before answering, where its predecessor answered directly. The training run behind that is large for a model of this class: Liquid reports about 33 trillion pre-training tokens, up from about 17 trillion for the predecessor, while the model card lists a 36 trillion-token overall training budget. It was finished with reinforcement learning carried out inside real agent harnesses — OpenClaw, Hermes Agent, Pi and OpenCode — rather than only on static tasks. That last choice explains the shape of the results below: this model was trained in the environment it is meant to work in.

Split the vendor table in two

Read in full, Liquid's table describes a specialist. On tool use it is strong. In Liquid's tests, LFM2.5 scores 27.69 on AutomationBench against 8.72 for the Qwen model Liquid compares it with — the outlier the launch will be quoted for. It also leads on ToolSandbox and on a live agent benchmark, LiveClaw, at 48.09 against 38.19, and it narrowly edges Qwen on BFCLv4 function calling, 65.01 to 65.00 — an effective tie rather than a meaningful advantage. It leads all three instruction-following tests in the set. Liquid's own formulation is that it leads or matches on five of the eight tool-use and agentic benchmarks in the table. For an agent whose job is to follow a procedure and call the right tool, those are the relevant rows.

On knowledge and coding, the same table reverses. GPQA Diamond sits at 74.42 against 83.69 for the larger Qwen model. SWE-Bench Verified, run inside OpenCode, is 51.00 against 66.20. WebArena Verified is 26.75 against 37.62, and SciCode is 32.71 against 37.35. Liquid does not hide this; its own post tells users to add retrieval or search for knowledge-heavy work. That advice is the honest summary of the release. The model has been optimised to act, not to know.

One more discount applies to the headline win. Three times the next score on AutomationBench still leaves the winning score at 27.69, and the banking tool benchmark it leads, τ³-Bench Banking, is led at 8.04 against Qwen's 7.63 — both low. These are low absolute numbers on hard agent tasks, for every model in the table. A local agent that completes roughly a quarter to a half of the work in these tests is a tool for supervised workflows, not a replacement for a person checking the output.

The practical catches

Two details decide whether this is usable rather than merely downloadable. The first is memory. Active parameters set the compute bill, but all 24 billion parameters have to live somewhere: total weight storage and runtime memory still depend on precision, quantization, context length and serving configuration, not on the 2B figure in the name. The second is the licence. The checkpoint is marked lfm1.0 on Hugging Face, referring to Liquid's LFM Open License v1.0, not Apache 2.0 or MIT. The licence allows commercial use only while the relevant legal entity remains below $10 million in annual revenue; at or above that threshold, commercial use requires separate permission from Liquid AI. 'Legal Entity' includes entities under common control, so a small subsidiary cannot assume only its own revenue counts. Qualifying nonprofits doing noncommercial or research work are exempt from the threshold. The licence permits modification and redistribution without requiring derivative models to be open-sourced, subject to licence and attribution conditions. Open weights you can download are not the same as open weights you can ship.

None of that cancels the release. A model that can be tested by anyone, on their own machine, against the vendor's published table is a better starting point than a hosted launch with a chart and no weights. The verification Liquid invites is the part worth doing. Run the tool benchmarks on your own tasks, and see whether the specialist split holds.

Sources

  1. [1] Liquid AI, "Run fast, private agents on-device with LFM2.5-24B-A2B", Hugging Face blog, 13 October 2026Read source
  2. [2] Liquid AI — LFM2.5-24B-A2B model card (Hugging Face)Read source
  3. [3] Liquid AI — LFM Open License v1.0 (full legal text)
  4. [4] Unite.AI, "Liquid AI Releases Open-Weight LFM2.5-24B-A2B for On-Device Agents", 13 October 2026Read source
  5. [5] DEV Community, "Liquid AI ships open-weight 24B agent model for local use", 13 October 2026Read source
  6. [6] VentureBeat, coverage of LFM2.5-2.6B licensing, 2026Read source