Liquid AI releases LFM3: six 2B-class open-weight models for phones and laptops, with a $10M revenue cap

Liquid AI has released LFM3, a family of small models meant to run on phones, laptops and other devices. The Liquid blog and Hugging Face post describe six models, all built from one base, LFM3-1.7B-Base:

• LFM3-1.7B: text chat.
• LFM3-VL-2B: vision and language.
• LFM3-Audio-2B: speech in and out.
• LFM3-Omni-2B: text, images and audio together.
• d1-2B and d1-Omni-2B: Liquid’s decision models, with d1-2B replacing the earlier d1-3B we covered in our d1 story.

What changed inside

Liquid says the base saw 22 trillion tokens of pre-training plus a 3.5 trillion token mid-training stage. The architecture is what Liquid calls a “tri-hybrid”: 24 layers mixing 13 short convolution layers, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. Liquid says this lets LFM3-1.7B “outperform our previous LFM2.5-2.6B while using 35% fewer parameters”, and that the text model “fits in about 1 GB at 4-bit precision”, with a 256K context window and fine-tuning on a single GPU.

Those are Liquid’s numbers. We have not run the models yet.

The license is the catch

The weights are on Hugging Face, but this is not Apache or MIT. The LFM Open License v1.0 says commercial rights “are conditioned upon You or Your Legal Entity not exceeding the Threshold”, and sets that threshold at “annual revenue of 10 million United States dollars”. Hobby use, research and small companies are fine. A larger company shipping LFM3 in a product needs to talk to Liquid.