Liquid AI launches LFM3: open-weight 2B multimodal family on one backbone
JosephJacks_ · x · 2026-10-28
Liquid AI announced LFM3, a new generation of Liquid Foundation Models for specialized intelligence: text, vision, audio, omni, and decision models on a shared new backbone, first released in the 2B range. Trained on 22T tokens with a redesigned hybrid architecture (via hardware-in-the-loop search across 14 device/backend combos), it targets agentic use with high capability per parameter and fast long-context inference. Base and post-trained weights are open on Hugging Face, with local fine-tuning supported.
More from Multimodal
- Fixing AI image drift: use the approved design as a reference to remove noise — LudovicCreator · 2026-10-28
- One bag, three brand identities: comparing AI variants on the same board — LudovicCreator · 2026-10-28
- One Firefly Board, one coffee brand: full identity workflow from moodboard to launch poster — LudovicCreator · 2026-10-28
- MiniMax H3 generates stunningly realistic in-car selfie, user calls it best work yet — Error4010 · 2026-10-28
- MOSS-Image ecosystem roundup: ComfyUI 0.43, QwenVL layout node, new style LoRAs — optimisticalish · 2026-10-28
- Community releases CC0 1980s Eurodisco LoRA for open-source music model YuE2 — Robonotes1760 · 2026-10-28