Liquid AI releases open-weight multimodal LFM3 models around 2B
Original title: Introducing LFM3: Every Modality for Specialized Intelligence
LFM3-1.7B-Base was pre-trained on 22T tokens, with a 256K context after mid-training.
Liquid AI announced LFM3, a family of Liquid Foundation Models around 2B parameters that share one backbone for text, vision, audio, omni, and decision models. The first release is open-weight and available on Hugging Face, the Liquid AI Playground, and the Liquid AI API. LFM3 uses a tri-hybrid mix of gated short convolutions, sliding-window attention, and global attention, and was designed for on-device and agentic use. LFM3-1.7B is a text model with a 256K-token context window, post-trained for agent harnesses including Hermes Agent, Pi, and OpenClaw.
Unconfirmed: Speed and quality claims are Liquid AI's own and are not independently verified in the text.
- Of 24 layers, 13 are convolutions, 5 are 256-token SWA, and 6 are global attention.
- The final 3.5T tokens were a mid-training stage that extends context to 256K.
- At 8K tokens, LFM3-1.7B is about 25% faster per parameter than LFM2-1.2B, Liquid AI says.
- Hardware search covered 14 device and backend combinations, including phones, a Raspberry Pi 5, and NPUs.
- Vendors
- Liquid AI
- AI models
- LFM3LFM3-1.7BLFM3-1.7B-Base
From the original article
13 min read at liquid.aiIntroducing LFM3: Every Modality for Specialized Intelligence Today, we’re excited to announce LFM3, our new generation of Liquid Foundation Models (LFMs) built for specialized intelligence. The LFM3 model family is multimodal by nature and brings text, vision, audio, omni, and decision models onto one new shared backbone. The first release of the LFM3 family covers models in the 2B parameter range.
By Philpax · liquid.ai
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.