RC RANDOM CHAOS

Liquid's LFM3: Multimodal Models for Edge Deployment

· via Hacker News

Original source

Introducing LFM3: Every Modality for Specialized Intelligence

Hacker News →

Liquid AI released LFM3, a new generation of Liquid Foundation Models designed for multimodal edge deployment. The first release includes six 2B-parameter models—text, vision, audio, omni, and two decision variants—derived from a shared backbone. The architecture combines gated short convolutions, sliding-window attention, and global attention across 24 layers. Pre-training used 22T tokens with expanded coverage of scientific papers, code, and mathematics.

The design prioritizes speed and efficiency. On-device testing shows the text model decoding 1.8 to 2.8 times faster than competitors while using less memory—a 128K-token context needs 1.6 GB of KV cache versus 4.3 GB for Qwen3.5-4B. The model reached this performance through four-stage post-training including 1.75T tokens of supervised fine-tuning. The vision model accepts video input and delivers performance comparable to models six times larger on visual benchmarks.

All models are open-weight, available via Hugging Face and Liquid’s API, and support local fine-tuning.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.