Overview
- Liquid AI launched a family of six open‑weight multimodal models that share a 1.7B‑base architecture and are designed to run and be customized on local hardware using a single GPU.
- Each model is roughly 2 billion parameters and includes specialized variants for text, vision‑language, audio, decision making, and an omni model that handles text, images, video, and audio together.
- The text model supports a 256,000‑token context window and the models use a claimed tri‑hybrid architecture—gated short convolutions, sliding‑window attention, and global attention—pretrained on about 22 trillion tokens.
- Liquid AI reports that its vision‑language model, LFM3‑VL‑2B, matches performance of models more than six times larger on vision tasks but these benchmark results come from the vendor and await independent testing and real‑world deployments.
- The open‑weight release aims to let businesses keep data on‑premises and adapt models for local use cases such as robotics and factory defect detection, and it follows Liquid AI’s recent October release of a larger MoE model that signals the company is pursuing both compact and large‑sparse approaches.