Liquid AI released LFM2.5-2.6B, a small open-weight model built to plan tasks and call external tools while running entirely on a laptop or phone rather than in the cloud. It fits in under 2.5 GB and generates 220 words-per-second-scale output on a high-end laptop processor, with day-one support for common local and server inference tools. The company notes that coding-heavy or more complex tasks are still better served by larger models.
What changed
Liquid AI's previous small models were less focused on tool-calling and multi-step agent work, and the earlier LFM2.5-8B-A1B used a much smaller instruction-tuning mix.
What it unlocks
Running an assistant that plans and calls tools entirely on a laptop or phone, with no per-token cloud cost and no data leaving the device.
- 2.6B parameters, under 2.5 GB memory
- 220 tokens/s on an M5 Max CPU, 113 on a Ryzen AI Max+ 395, 30 on a phone
- pre-trained on ~34T tokens, 128K context, 128K vocabulary
- ~15K output tokens/s at high concurrency on one H100, roughly 1.3B tokens per day
What you need to act on it
- downloading the open weights from Hugging Face
- serving the model locally behind an OpenAI-compatible endpoint and connecting an agent harness
- liquid.ai2026-08-04