LiquidAI releases LFM3, a family of 2B-parameter multimodal foundation models on Hugging Face
LFM3 offers a shared backbone for text, vision, audio, and decision models, with up to 35% fewer parameters than its predecessor and support for a 256K context window.
1 source · cross-referenced
- LiquidAI released LFM3, a family of six 2B-parameter multimodal models sharing a single backbone for text, vision, audio, and decision tasks.
LiquidAI published LFM3, a new generation of Liquid Foundation Models (LFMs) on Hugging Face, comprising six 2B-parameter models that share a single backbone for text, vision, audio, and decision tasks.
The LFM3 family includes LFM3-1.7B (text), LFM3-VL-2B (vision-language), LFM3-Audio-2B, LFM3-Omni-2B (multimodal), and two decision-focused variants, d1-2B and d1-omni-2B.
LFM3-1.7B-Base uses a tri-hybrid architecture with 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers, and was pre-trained on 22T tokens before a mid-training stage on 3.5T curated tokens that extended the context window to 256K.
LiquidAI reports LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters, citing a 38.42 score on WebArena-Verified against 3.02 for its predecessor, and advantages on agentic tasks in browsers and business apps.
The text model fits in about 1 GB at 4-bit precision, and the vision and audio encoders are designed for edge deployment, with the audio encoder enabling time-to-first-token as low as 50 ms and LiquidCodec supporting streaming decoding at about 87x real time on an Apple M4 Max CPU thread.
LFM3 models are available via Hugging Face Transformers, with additional inference engines listed on their model cards, and can be fine-tuned on a single GPU for domain-specific adaptation.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79