Developers building private, offline features on devices get small free models, though commercial use is limited to firms under $10 million revenue.
Kenning's take
What it does
LFM2.5 is a family of 1.2 billion parameter models from Liquid AI designed to run directly on devices rather than in the cloud. The release includes base and instruction-tuned text models, a Japanese model, a vision-language model and an audio-language model that Liquid says is 8 times faster than its predecessor. The weights are on Hugging Face and Liquid's LEAP platform under the LFM Open License, which allows free commercial use for companies under $10 million in annual revenue.
- 1.2B parameter Base and Instruct text models
- Japanese, vision-language and audio-language variants
- Pretraining extended from 10T to 28T tokens
- Tuned for instruction following and tool use
- NPU optimisation with launch partners AMD and Nexa AI
Who it suits. Developers building AI features that must run privately and offline on devices.
Key facts
- Category
- AI models
- Made by
- Liquid AI
- Released
- 5 January 2026
- How to get it
- open weights on Hugging Face and Liquid's LEAP platform; Liquid also offers an API
- Works on
- self-hosted, Hugging Face, on-device
- In the UK
- Yes
- Free plan
- free weights for research, non-commercial use, and commercial use under $10M annual revenue
FAQ
When was LFM2.5 released?
Liquid AI announced the LFM2.5-1.2B model family on 5 January 2026.
Is LFM2.5 free for commercial use?
Under the LFM Open License v1.0, commercial use is licensed only for organisations with less than $10 million in annual revenue.
What models are in LFM2.5?
Base and Instruct text models, a Japanese model, a vision-language model and an audio-language model, all at the 1.2B size.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
ACE-Step 1.5
Open-source music generation model that writes full songs with lyrics and runs on ordinary consumer GPUs.
AlphaGenome Atlas
Google DeepMind database of AI-predicted effects for every possible single-letter change in human DNA.
Amazon Nova 2
Amazon's second family of Nova AI models, covering reasoning, speech conversation and combined text-and-image generation.
Cisco Antares
Cisco's small open-weight models that find where known security vulnerabilities sit in a codebase, running locally.