Liquid AI’s tiny 1B

Liquid AI, the edge AI startup, has open-sourced a new model that could be the perfect brain for your next phone — and Apple might be eyeing the company itself. According to The Information, Apple has Liquid AI on its acquisition shortlist as it pushes to bring more intelligence on-device with iOS.

The model, called LFM2.5-8B-A1B, is a mixture-of-experts design: 8 billion total parameters, but it only activates 1 billion per token. That means it can run locally on a phone or laptop without breaking a sweat. It's released under the LFM open weights license, so you can fine-tune it on a single consumer GPU and run it fully offline.

This isn't your typical one-shot generator. LFM2.5-8B-A1B is a reasoning model that shows its chain of thought before answering — think of it as a tiny chatbot that stops to think first. And because it only uses 1B active parameters, that thinking doesn't drag down your device. The context window has jumped from 32K to 128K tokens, enough to swallow a whole book or a hefty codebase. The vocabulary was doubled to 128K tokens too, making it much better at non-English languages: Thai and Hindi processing speeds improved by 238.2% and 120.4% respectively, and Arabic got a 38.8% boost.

Liquid's team also tackled two common problems with small local models: getting stuck in infinite loops and hallucinating. They optimized the architecture to avoid the mental merry-go-round, and added targeted reinforcement learning on safety guardrails so the model will honestly say "I don't know" when it's out of its depth. The result? The no-hallucination rate jumped from a paltry 7.46% in the previous generation to 63.47%. On complex tasks and tool-use benchmarks, it handily beats the competition.

On day one, LFM2.5-8B-A1B works with llama.cpp and MLX, so you can run it locally right now. On an Apple M5 Max chip, it hits 253 tokens per second — and it even runs smoothly on a phone.