Liquid AI released LFM2.5-2.6B on August 4, a 2.6 billion parameter model built to plan, pick tools and finish multi-step jobs while running entirely on a phone, a laptop CPU or a single GPU. The company, which AI Insiders has covered before for its non-transformer LFM architecture, is arguing that agentic competence does not require frontier scale. That is a harder claim than it looks, because a chatbot’s single wrong answer stays contained while an agent’s early mistake propagates through every step that follows.
Liquid AI pretrained the base model on roughly 34 trillion tokens, then ran it through four post-training stages: supervised fine-tuning, teacher specialization by domain, on-policy distillation from those domain teachers, and a final reinforcement-learning stage conducted inside real agent harnesses including Hermes Agent and OpenClaw. Training directly inside the sandboxes an agent will actually operate in, rather than on static examples of agent behavior, is the mechanism Liquid AI credits for the scores below.
The company tested the model against four larger open-weight systems it chose itself: Google’s Gemma 4 in 5.1B and 8B configurations, and Alibaba’s Qwen3.5 in 4.7B and 9.7B configurations. On instruction following, LFM2.5-2.6B leads all four, scoring 59.17 on IFBench and 85.49 on IFStruct, both comfortably ahead of every rival. On the two dedicated function-calling benchmarks, BFCLv4 and ToolSandbox, it also leads or nearly leads, trailing only the 9.7B Qwen3.5-9B on BFCLv4, 56.88 against 60.13.
That is tool-calling accuracy: correctly identifying which function to invoke and how to format the call. Full agentic task completion, measured on harness-based benchmarks like Claw-Eval, PinchBench and BrowseComp+, tells a less flattering story. There, LFM2.5-2.6B stays within a few points of the Qwen models without pulling ahead of either one. It trails Qwen3.5-4B and Qwen3.5-9B on PinchBench, scoring 68.22 against their 71.26 and 71.45, and trails Qwen3.5-9B on BrowseComp+, 26.89 against 27.23. Coding shows the same pattern: on LiveCodeBenchv6, Qwen3.5-9B outscores it by more than ten points. Liquid AI’s own summary is candid on this point: the model beats systems nearly four times its size on tool selection and instruction following, but larger models keep the edge once a task runs long enough to require sustained reasoning across steps.
Liquid AI picked every comparison model in that table, a standard practice for a model launch but a reminder that these numbers describe performance against a self-selected field, not an independent leaderboard.
The speed figures are specific about which hardware produced them, and that specificity matters because “runs on a phone” can mean a current flagship or a five-year-old handset. Liquid AI measured 220 tokens per second on Apple’s M5 Max chip and 113 tokens per second on AMD’s Ryzen AI Max+ 395, both high-end laptop or desktop silicon rather than mobile processors, while the model’s memory footprint stayed under 2.5 GB. On an unspecified phone, the company reported 30 tokens per second, enough to keep an agent responsive, but without naming the device or chipset behind that figure. On a single Nvidia H100 GPU running SGLang, throughput climbed to nearly 15,000 tokens generated each second once request load was pushed to its highest tested level, which Liquid AI translates to roughly 1.3 billion tokens a day from one card.
Liquid AI leads with free inference, the removal of per-token API costs, as the headline benefit of running locally. That is the weaker half of the pitch. The stronger argument is that data processed by an on-device agent never leaves the handset: no network call carries a user’s messages, health data or location to a third-party server for inference. For any agent handling that category of information, local processing is a different privacy posture than a cheaper version of a cloud call, not merely a discount on the same architecture.
Both the base and post-trained checkpoints are live on Hugging Face today under what Liquid AI describes as an open-weight license permitting download, fine-tuning and deployment without restriction, with day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX. Teams building privacy-sensitive on-device agents, for messaging, health or field operations, get a model that already clears the tool-calling bar. Before committing production workloads to it over a 9B-class alternative, they should run their own multi-step task evaluations: Liquid AI’s own numbers show the model closing the gap on tool selection but not yet matching larger models on the harder job of finishing what it starts.
Liquid AI published LFM2.5-2.6B’s architecture, training pipeline and benchmark results in a company blog post on August 4, 2026.
