Editorial illustration for Liquid AI's LFM2.5-2.6B Runs on Raspberry Pi, Hits 15,000 Tokens/Second
Liquid AI's LFM2.5-2.6B Runs on Raspberry Pi, Hits...
Liquid AI, the startup founded in 2023 by a group of former MIT researchers, released a new open-weight language model this week called LFM2.5-2.6B. The name is a mash-up of the generation number (2.5) and the parameter count (2.6 billion), and it's built for a narrower job than most large models chase: running agentic tasks directly on local hardware, with no GPU and no cloud connection required. Liquid says the model works on everything from smartphones and laptops down to a Raspberry Pi, which opens up edge AI for use cases like tool calling, document handling, calendar automation, and background routines in vehicles or robotics where connectivity can't be guaranteed.
The model carries a 128,000-token context window and native tool-calling support, and Liquid has published both the post-trained version and a base checkpoint, LFM2.5-2.6B-Base, on Hugging Face for developers who want to fine-tune it. In an interview with VentureBeat, Liquid's researchers laid out where the model fits, and where it doesn't, compared to running larger systems in the cloud.
In release materials and a recent interview with VentureBeat, Liquid's researchers said LFM2.5-2.6B can run entirely on local hardware — from smartphones and laptops down to a Raspberry Pi — without relying on cloud inference or GPUs, unlocking edge AI applications and giving more options to enterprises working in regulated industries or with sensitive information they don't want to send up to the cloud.
Why this matters
The range Liquid AI is claiming here is the story: a 2.6-billion-parameter model that runs on a Raspberry Pi and, on an H100, pushes close to 15,000 tokens per second under load. That's roughly 1.3 billion tokens a day from a single GPU, no cloud round-trip required. For developers building agents, that math changes the cost conversation.
If LFM2.5-2.6B holds up outside Liquid's own benchmarks, teams that have been paying per-token for hosted inference get a real alternative for latency-sensitive or offline work, drones, kiosks, factory sensors, anything that can't wait on a network call. The Apollo app lowers the barrier further; anyone can test it on a phone before committing engineering time. We'd still want independent numbers on accuracy and failure modes before treating this as a drop-in replacement for larger hosted models.
Liquid, founded by former MIT researchers in 2023, is betting that "runs anywhere" beats "runs best." Worth watching whether enterprises actually deploy it on constrained hardware or just use the Pi demo as a headline.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives OpenAI's AI Agents Used Message Board to Plan Hacking Spree
From Archives OpenAI Browser Flaw Could Spam WhatsApp Contacts Via Prompt Injection
Further Reading
- Liquid AI’s LFM2.5-2.6B brings powerful on-device agents to Raspberry Pi-class hardware - VentureBeat
- Deploy local agents everywhere with LFM2.5-2.6B - Hugging Face Blog
- LFM2.5-2.6B: Deploy Agents Everywhere - Liquid AI Blog
- Running Liquid AI's New Model LFM2.5-2.6B on Raspberry Pi 5 - note.com
- LFM2 Technical Report - arXiv
