Liquid AI Ships LFM2.5-2.6B, a Phone-Ready Agent Model

The interesting bit in Liquid AI's release of LFM2.5-2.6B is not the parameter count, it is where the parameter count is supposed to land. A 2.69B model that fits in under 2.5 GB of memory, runs at roughly 30 tokens per second on phone-class hardware, and, per the model card's own benchmark table, beats or matches models several times its size on instruction following and tool use.

The architecture is the tell about what Liquid is optimizing for. Thirty layers split into 22 double-gated short convolution blocks and 8 grouped-query attention layers, a 128K context window, and around 34 trillion training tokens. On an Apple M5 Max the company reports 220 tokens per second, and 113 tokens per second on an AMD Ryzen AI Max+ 395. On a single H100 at high concurrency the model card claims roughly 15K output tokens per second, or about 1.3B tokens per day. The post-training pipeline runs agentic reinforcement learning inside popular harnesses like AutoGen and LangGraph, which is unusually specific for a model card and hints at where the team thinks the practical wins live.

Why any of this matters if you are not shopping for a small model this week: on-device agents have been the promised land for a while, but the honest history is that small open models kept losing to hosted systems on the messy multi-step stuff. A credible 2.6B agent that runs on a laptop CPU changes the cost math for anyone building assistants against private data, offline apps, or workloads where the round-trip to a cloud model is too slow or too expensive to justify. Liquid is already a recurring name in this band, including their recent partnership with MacPaw for on-device Mac inference.

The numbers to hold at arm's length: the benchmarks are the company's own selection against a specific set of comparators (gemma-4 and Qwen3.5 variants), and Liquid itself notes the model is not recommended for agentic coding or knowledge-heavy tasks. The card does not tie the phone number to a specific handset, and the model ships under Liquid's own lfm1.0 license rather than a standard permissive one, so anyone with commercial plans should read the terms before building on it. Treat the throughput and win-rate figures as vendor claims until an independent run lands.

If the tool-use numbers hold up outside the leaderboard, the downstream story worth watching is who else races to ship a genuinely useful sub-3B agent, because that is the size band where the difference between a cool demo and something that runs on the fleet of laptops a company already owns gets decided.