A Full AI Agent Now Fits in 2.5 GB Right on Your Phone

5 August 2026 • Lucas Ferretti • 2 min read

2.5 GB max memory required

220 tok/s speed on Apple M5 Max

30 tok/s speed on a smartphone

34 T pre-training tokens

2.5 GB of memory. That is all it now takes to run an autonomous AI agent on a smartphone. Liquid AI just released LFM2.5-2.6B, and the cloud is officially optional.

LFM2.5-2.6B, an AI agent sized for your pocket

Released on August 4, this 2.6-billion-parameter model plans, calls tools, and chains multi-step tasks on local hardware. The open weights sit on Hugging Face, in base and post-trained versions.

Speed is the showstopper. Liquid AI reports 220 tokens per second on an Apple M5 Max, 113 on a Ryzen AI Max+ 395, and roughly 30 on a phone. One developer already got an agent demo to run fully inside a browser.

Why LFM2.5-2.6B rivals models four times its size

Pre-training consumed close to 34 trillion tokens. On tool use, the model even edges past Qwen3.5-9B on the ToolSandbox benchmark, 77.83 points to 76.44.

The recipe rests on four post-training stages, with reinforcement learning run inside real agent harnesses such as Hermes Agent, OpenClaw, and Pi. Smart play: win the one category that matters on-device rather than every leaderboard.

What this changes for you and for developers

A local agent sends zero data to a server, costs zero per token, and answers without network lag. Your emails, files, and requests stay on the device you own.

Developer support lands on day one: llama.cpp, MLX, vLLM, SGLang, and ONNX. On a server, a single H100 pushes nearly 15,000 tokens per second, enough to unleash fleets of agents at near-zero marginal cost.

Honest caveats remain. Code generation is still the turf of much bigger models, and the in-house lfm1.0 license deserves a careful read before commercial use. While open-weight giants like Kimi K3 bet on sheer scale, Liquid AI bets on the smallest useful model.

Is LFM2.5-2.6B free to download?

Yes, both the base and post-trained versions are free on Hugging Face under Liquid AI's lfm1.0 license. Commercial projects should review its terms first.