Liquid AI's 2.6B Model Runs Agents on Your Phone
LFM2.5-2.6B punches well above its weight class — matching models four times larger while fitting in under 2.5 GB of RAM.
By The ai-compute Desk
Most edge AI stories follow the same arc: a company shrinks a model, something important breaks, and developers quietly reach for the cloud anyway. Liquid AI's LFM2.5-2.6B is trying to rewrite that arc. [1]
The headline number is memory: under 2.5 GB. [1] That figure puts the model inside the envelope of a mid-range smartphone, an AMD Ryzen AI Max+ 395, or an Apple M5 Max — no cloud subscription required. On that M5 Max, Liquid AI clocks decode speeds of 220 tokens per second; on a Ryzen AI Max+ 395, the figure is 113 tokens per second. [1]
Speed at the edge is only half the pitch. The harder problem is making a small model actually useful as an agent — able to call tools, follow instructions across multiple turns, and adapt to different system prompts without hallucinating a tool name or losing the thread. Liquid AI's answer is a four-stage post-training pipeline. First, two rounds of supervised fine-tuning weighted toward agentic data — tool use, web search, and harness trajectories. Then teacher specialization, where one specialist teacher per domain (math, code, tool use, and more) is trained. This is followed by a separate stage Liquid AI calls Multi-domain on-policy distillation (MOPD), which distills the specialist teachers into a single student. [1]
The final stage is the most distinctive. Liquid AI runs multi-turn reinforcement learning inside real agent harnesses, where the model learns to work across different tools, system prompts, and multi-turn task environments. [1] A Harness Proxy lets the team treat agentic harnesses as black boxes with no modification, while transparently capturing the token-level trajectories needed to reconstruct and validate RL training samples. [1] The model learns to navigate actual tool-calling loops, not simulations of them.
On benchmarks, LFM2.5-2.6B tops every instruction-following benchmark in its comparison group and every tool-use benchmark except BFCLv4, where only the 9.7B Qwen edges ahead. [1] Evaluated against models up to roughly four times its size, on agentic tasks it beats both Gemma models tested and stays even with the Qwens. [1] Coding is the acknowledged weak spot: larger models keep a clear lead there, and Liquid AI says as much plainly in its documentation. [1]
GPU throughput tells a complementary story for server-side deployments. At high concurrency, LFM2.5-2.6B reaches almost 15,000 output tokens per second on a single H100 — roughly 1.3 billion tokens per day. [1] The model ships with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX, which means it slots into existing inference stacks without a rewrite. [1] A WebGPU browser demo is also live, requiring no local setup.
The logic underneath all of this is structural, not just technical. When a capable agent can run locally, it changes the economics of who can deploy AI and where — removing latency, licensing costs, and the privacy tradeoffs that come with routing sensitive data through external servers. The edge is not a consolation prize for teams who cannot afford the cloud; it is increasingly a design choice.
LFM2.5-2.6B and its base variant are both available on Hugging Face today, and Liquid AI's documentation includes a guide for running local agents using harnesses like OpenClaw, Hermes Agent, and Pi. [1] Context window extends to 128K tokens — added through a mid-training phase that expands the range of tasks the model can handle without chunking. [1]
The gap between cloud-scale agents and on-device agents is narrowing faster than the industry expected. LFM2.5-2.6B won't replace a frontier model for complex code generation, but for the broader category of tool-calling, instruction-following agents — the workhorses of most production deployments — a 2.6B model that fits on a phone is a credible option. [1] That changes the conversation.
At 30 tokens/s, it allows you to run capable agents even on a phone.Liquid AI