Liquid AI ships LFM2.5-2.6B agent model that runs on-device

The interesting thing about Liquid AI's release of LFM2.5-2.6B is not the leaderboard chart, it is where the model is meant to run. A 2.6 billion parameter model, small enough to sit under 2.5 GB of memory, aimed squarely at agent workloads such as planning, tool calling, and multi-step task completion, on the device the user already owns.

Liquid AI's own numbers put it at 220 tokens per second on an Apple M5 Max, 113 on an AMD Ryzen AI Max+ 395, and 30 on a smartphone. On a single NVIDIA H100 they report nearly 15,000 output tokens per second at high concurrency, or roughly 1.3 billion tokens per day from one card. The company's framing is direct: "Local agents give you free inference, low latency, and real privacy. Removing the per-token cost changes how developers build."

The economic angle is the one worth pausing on. Frontier agent products today burn through millions of tokens per session, and that meter runs against a paid API. If a 2.6B model can carry a real slice of the background work, the planning loops and the tool-calling that never surfaces to the user, the marginal cost of that work goes to zero. Liquid describes agents that "can now be massively parallelized on local hardware, running background tasks that burn through millions of tokens at no marginal cost." That is a different unit economics than a hosted API bill.

The honest caveat is that these benchmark numbers are self-published, and instruction-following scores like Multi-IF at 80.07 or ToolSandbox at 77.83 do not prove reliability on the odd, novel task a real user actually hands an agent. What the blog post does not give you is battery impact on the phone, commercial licensing terms for the Hugging Face release, or how the model behaves outside the harnesses Liquid tested. Take the specifics as reported, not as settled.

If it holds up, the winners are the developers building agentic tooling whose inference bill is currently the whole business, and the privacy-sensitive verticals such as health, legal, and defense that could never send tokens to a cloud endpoint in the first place. The direction, small models tuned specifically for agentic loops on the device, is the part worth watching.