Deploy local agents everywhere with LFM2.5-2.6B
Hugging Face Blog
Liquid AI dropped LFM2.5-2.6B, a tiny model built to run AI agents right on your laptop or phone. It matches models 4x its size on tool use, and runs fast enough for real-time agents without a cloud bill.
Liquid AI just shipped LFM2.5-2.6B, a 2.6-billion-parameter model that's less about raw smarts and more about where it can live. The pitch is simple: run a genuinely capable agent on a phone or laptop, with no server round-trip, no per-token cloud invoice, and no data leaving the device. On an Apple M5 Max it decodes at 220 tokens per second. On an AMD Ryzen AI Max+ 395 CPU, 113 tokens per second. Even on phone-class hardware, Liquid says it holds a usable 30 tokens per second, all while staying under 2.5GB of memory.
What's notable is how deliberately this thing was built for agentic work rather than general chat. The training pipeline runs through four stages after a 34-trillion-token pretrain: heavy supervised fine-tuning on tool-use and web-search trajectories, specialist teacher models for domains like math and code, a distillation step that folds those specialists into one student model, and finally reinforcement learning conducted inside actual agent harnesses like OpenClaw and Hermes Agent. That last piece is the interesting bit — the RL loop separates training, inference, and sandboxed task execution into distinct services, letting the model learn from real multi-turn tool interactions rather than static datasets.
The benchmark numbers back up the framing. Against models nearly 4x its size — Qwen3.5-9B, Gemma-4-E4B — LFM2.5-2.6B tops every instruction-following test in the comparison and wins nearly every tool-use benchmark, losing only to the 9.7B Qwen on BFCLv4. It also edges out both Gemma variants on agentic benchmarks like ToolSandbox and BrowseComp+, and holds its own against the larger Qwens. Coding is the exception: LiveCodeBench scores trail the bigger models, so Liquid isn't pretending this is a code-generation champion. It's an agent-execution specialist that happens to be small.
On the GPU side, the story shifts from mobile efficiency to throughput. A single H100 pushes nearly 15,000 output tokens per second at high concurrency, which Liquid translates into roughly 1.3 billion tokens a day from one card. Combined with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX, the model is clearly aimed at developers who want to deploy the same weights from a phone app to a server rack without re-architecting anything. Both LFM2.5-2.6B and its base checkpoint are live on Hugging Face now, alongside a browser-based WebGPU demo.
My take
Small, agent-tuned models like this are the more honest bet in AI right now — most real-world tasks don't need a frontier-scale brain, they need reliable tool calls and privacy, and Liquid is chasing exactly that instead of another leaderboard flex. The RL-inside-real-harnesses approach is the smart part here, and it's a pattern more labs should copy instead of training on static datasets and hoping it generalizes to messy agent loops.
Read more about this at: Hugging Face Blog
