On this page
Liquid AI’s blog post dated October 27, 2026—titled Introducing LFM3: Every Modality for Specialized Intelligence—announces LFM3, a new generation of Liquid Foundation Models. The family is multimodal by design and puts text, vision, audio, omni, and decision models on one shared backbone. The first release covers models in the about 2B parameter range. Base and post-trained weights are open-weight and available on Hugging Face, Liquid Playground, and the Liquid AI API. Related open-weights reading: Atria Dawn Formal Weights Are Out: 512K Context on Hugging Face.
Architecture and training (vendor description)
Liquid describes a new tri-hybrid stack: gated short convolutions, sliding-window attention (SWA), and global attention. On the Hugging Face card for LFM3-1.7B, the text model is about 1.73B parameters, with 24 layers (13 convolution / 5 SWA with a 256-token window / 6 global attention), a 256K context window, a 128K vocabulary, and the LFM Open License v1.0. Liquid says LFM3-1.7B-Base was pre-trained on about 22T tokens, with a final mid-training stage of about 3.5T tokens that also extends text context to 256K.
The blog states that LFM3-1.7B is more expressive than the prior LFM2.5-2.6B while using about 35% fewer parameters. Benchmark tables and on-device throughput figures on the post are Liquid’s own reported numbers—for reference only.
Six about-2B-class models from one backbone
The first LFM3 release lists six about-2B-class models branched from LFM3-1.7B-Base (quantized siblings on Hugging Face are packaging variants, not separate product lines):
- Text — LFM3-1.7B (256K context)
- Vision — LFM3-VL-2B (including video input)
- Audio — LFM3-Audio-2B (text or audio in; text and/or speech out; Liquid says support for about 18 languages)
- Omni — LFM3-Omni-2B (vision + audio + language)
- Decision — d1-2B (from VL) and d1-Omni-2B (from Omni), aimed at millisecond multi-question decisions with calibrated probabilities
Related open-weights reading: MiniMax M3.1 Flash Open Weights Are on Hugging Face. For speech / multimodal edge reading: Block Releases Dessa, a Full-Duplex Speech Model Trained Only on Speech.
Day-0 partners and where weights are today
Launch partners named on the blog include Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm (AI Hub / GenieX), and Vercel. Builder ecosystem mentions include Hugging Face, llama.cpp, MLX, Ollama, and vLLM, among others.
OpenRouter appears on that day-0 partner list. OpenRouter’s public models list still shows Liquid entries in the lfm-2.5-* line only—**no liquid/lfm3-* (or equivalent) model IDs. Weights and hosted access today are on Hugging Face, Liquid Playground, and the Liquid AI API**.
Availability, fine-tuning notebooks, and local run paths follow Liquid’s LFM3 blog and the LFM3-1.7B model card. For reference only.