Liquid AI's new 2.6B model runs AI agents at 220 tok/s on a MacBook

Liquid AI just dropped a small model that punches far above its weight class, and the numbers are frankly a little absurd. Their new LFM2.5-2.6B, a 2.6-billion-parameter model, is explicitly designed to run autonomous AI agents locally on everything from a phone to a server. The headline stat? It cranks out 220 tokens per second on an Apple M5 Max and 113 tok/s on an AMD Ryzen AI Max+ 395 CPU, all while sipping under 2.5 GB of memory. That’s fast enough to feel instantaneous, and it’s a genuine unlock for on-device agentic workloads that previously needed a cloud GPU.

But raw speed is just the cost of entry. What’s more interesting is how they trained it. The post-training pipeline has four distinct stages, and the final one, called Agentic Reinforcement Learning, is the secret sauce. Instead of just optimizing for benchmark scores, Liquid AI ran multi-turn RL directly inside real agent harnesses like OpenClaw and Hermes Agent. They built a sandbox service that treats these harnesses as black boxes, letting the model learn to use tools, follow system prompts, and navigate multi-step tasks in the same environment you’d actually deploy it in. The RL framework captures token-level trajectories for training without modifying the harness itself. It’s a clever bit of infrastructure engineering that bridges the gap between a good language model and a reliable agent.

The benchmark results are where the smirk really sets in. LFM2.5-2.6B tops every instruction-following and tool-use benchmark they tested against models up to 4x its size, including the 9.7B Qwen. On agentic tasks, it beats both Gemma models and stays even with the Qwens. The one honest admission: coding is still a weak spot. Larger models keep a clear lead there, so Liquid AI suggests reaching for something bigger if you’re generating production code.

This feels like a direct shot at the “giant model or bust” mentality that’s obsessed the industry for two years. A 2.6B model that runs agents on a phone and trades blows with models four times its size is a reality check. The practical upshot is that high-volume, privacy-sensitive agent workloads—think on-device research assistants, local data analysis, or personal automation—suddenly look a lot more feasible without a constant internet connection. Both the instruct and base models are available on Hugging Face right now, with day-one support for llama.cpp, MLX, vLLM, and ONNX. The era of the local agent might finally be getting its engine.