- Published on
Liquid AI's new model LFM2.5-2.6B brings powerful AI agents to devices as small as a Raspberry Pi
- Authors
- Name
- DeepHorus
Liquid AI, a 2023 MIT spinoff, unveiled LFM2.5-2.6B this week, an open-weight model for agentic tasks that runs on local hardware. It works on devices from smartphones to a Raspberry Pi, bypassing cloud inference and GPUs. The company targets edge AI and enterprises with sensitive data or strict regulations.
The startup was founded by former MIT computer scientists who argue that a sufficiently capable small model can handle well-defined agentic workloads locally, such as tool calling, document management, calendar automation, and background routines. It suits connectivity-limited environments like vehicles and robotics, while coding-heavy tasks remain best on larger frontier models.
LFM2.5-2.6B packs 2.6 billion parameters, a 128,000-token context window, and native tool calling. Both the tuned model and a base checkpoint are available on Hugging Face, with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. An open-source fine-tuning framework, LEAP, is also released.
The model's LFM2 architecture is explicitly designed for real-world CPU performance rather than GPU benchmarks. Company-reported measurements show 220 tokens per second on an Apple M5 Max and 113 on an AMD Ryzen AI Max+ 395, using less than 2.5 GB of memory. A smartphone reaches about 30 tokens per second, while a single Nvidia H100 sustains nearly 15,000 output tokens per second under load. These figures are vendor benchmarks, not independently verified.
For enterprises, the appeal is running task-specific agents essentially for electricity cost, avoiding per-token cloud fees. Regulated industries like healthcare, finance, and government can keep sensitive data on premises. Field scenarios such as connected vehicles and robotics gain from low latency and no network dependency.
"We want to make models for another type of user," Maxime Labonne, head of post-training, told VentureBeat. He said cloud models still have their place, but edge AI is key when cloud use is impossible. Labonne cited Raspberry Pi demos running "pretty fast," emphasizing the architecture's CPU-first design.
Users can try the models through Liquid AI's Apollo mobile app, and developers can fine-tune using the base checkpoint and LEAP. Liquid positions the release not as a rival to frontier models, but as a practical option for latency-sensitive, private, or cost-driven deployments.
The custom open weights license, like Moonshot's Kimi K3, warrants careful legal review. Liquid's launch reflects a broader industry shift toward hybrid AI, pairing cloud powerhouse models with on-device agents. This may accelerate adoption by making local AI a default for routine, high-volume enterprise tasks.
