Liquid AI's new 2.6B model runs AI agents directly on your laptop or phone
LFM2.5-2.6B is a compact AI model that can handle multi-step tasks without sending your data to the cloud, and it competes with models four times its size on key benchmarks.
Key points
- Liquid AI released LFM2.5-2.6B in August 2026, a 2.6-billion-parameter model designed to run AI agents entirely on a local device.
- The model reaches 220 tokens per second on an Apple M5 Max chip and 113 tokens per second on an AMD Ryzen AI Max+ 395 CPU, fitting in under 2.5 GB of memory.
- On instruction-following benchmarks, LFM2.5-2.6B outscored every comparison model tested, including models nearly four times larger.
- Both the base and instruction-tuned versions are available now on Hugging Face.
Most AI assistants that handle complex tasks live in the cloud. Your request travels to a remote server, gets processed, and returns. That works fine until privacy, cost, or internet access becomes a concern.
Liquid AI wants to change that. The company has released LFM2.5-2.6B, a small model built to run AI agents, software that carries out multi-step tasks on its own, directly on everyday hardware. Think the laptop on your desk or the phone in your pocket.
What makes this model different from a standard chatbot?
Standard chatbots answer one question at a time. An AI agent plans and acts across several steps: it might search for information, call an external tool, check a result, then write a summary. That demands precise instruction-following across a longer chain of decisions.
LFM2.5-2.6B was trained specifically for that kind of work. Liquid AI put it through four training stages, finishing with Agentic Reinforcement Learning, where the model practiced tasks inside real agent environments rather than reading text examples. It topped every instruction-following benchmark in Liquid's comparison group, beating models with up to 9.7 billion parameters on several tests. We've been watching this model family since Liquid AI's encoder release on 28 July, which showed similar efficiency gains on long documents.
How does the size comparison actually play out?
The table below shows LFM2.5-2.6B against four larger models on a selection of Liquid's published benchmarks. Higher is better on all rows except AA Omniscience, where scores are negative and closer to zero is better.
| Benchmark | LFM2.5-2.6B | Qwen3.5-4B (4.7B) | Qwen3.5-9B (9.7B) |
|---|---|---|---|
| IFBench (instruction following) | 59.17 | 48.40 | 56.47 |
| ToolSandbox (tool use) | 77.83 | 75.55 | 76.44 |
| Multi-IF (multi-turn instructions) | 80.07 | 55.67 | 62.55 |
| AIME25 (maths) | 51.87 | 49.33 | 56.07 |
| LiveCodeBench v6 (coding) | 59.41 | 60.85 | 69.86 |
Instruction following and tool use go to the smaller model. Coding is where the larger models hold a genuine lead, so developers building code-heavy agents will want something bigger.
Should ordinary users care about any of this?
Yes, for one practical reason: privacy. When an AI agent runs on your device, your documents and queries never leave it. No cloud company stores or trains on your data, which matters if you're handling sensitive personal or business information.
Speed is also meaningful in daily use. At 30 tokens per second the model's quick enough to run on a modern phone, not just a high-end laptop. It fits in under 2.5 GB of memory, well within reach of current consumer devices.
For developers, the model's available now on Hugging Face and works with widely used tools including llama.cpp and vLLM, a framework for serving AI models at scale.
Common questions
Does running AI on my device mean my data stays private?
Yes. When a model runs locally, your input never travels to an external server, so no third party receives or stores it. That's the core privacy benefit of on-device AI.
Do I need expensive hardware to run this?
Not necessarily. Liquid says the model runs on a phone at useful speeds and fits in under 2.5 GB of memory, less than many photo-editing apps use. A modern laptop or high-end smartphone should be sufficient.
