Liquid AI Releases the LFM3 Model Family: Six 2B-Class Open-Weight Models on One Backbone

What happened

The new architecture is the technical starting point of this release: LFM3 uses a tri-hybrid of gated short convolutions, sliding-window attention and global attention, with 13 convolution layers, 5 sliding-window attention layers with a 256-token window and 6 global attention layers across 24 layers, and the architecture search covered 14 device and backend combinations, including the Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI and the Qualcomm Hexagon NPU.

The six models divide labor as follows: LFM3-1.7B handles text with a 256K-token context window; LFM3-VL-2B handles text, images and video; LFM3-Audio-2B takes text or audio input and outputs text or speech, supporting 18 languages; LFM3-Omni-2B, the company's first omni model, accepts text, images, video and audio and outputs text and speech; d1-2B and d1-Omni-2B are decision models that return calibrated-probability answers in a single forward pass, adding four typed answer forms: point, read, estimate and pick.

The release terms and positioning were announced together: the base and post-trained models are available as open weights on Hugging Face, as well as on the Liquid AI Playground and the Liquid AI API, with day-0 launch partners including Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm and Vercel, and Mercedes-Benz and Shopify joining pre-release testing and evaluation. Company-reported benchmark figures include 71.39 on BFCL V4 tool calling and 96.66 on IFEval for the text model, and a 71.4 average across 62 vision benchmarks for the vision model; all figures are company-reported, with no independent reproduction yet.

Key facts

1. Release: Liquid AI published LFM3 on its official blog on Oct 27, positioned as a new generation of Liquid Foundation Models for specialized intelligence.
2. Scale: the first six models are all in the 2B parameter range, sharing the LFM3-1.7B-Base backbone, pre-trained on 22T tokens.
3. Text model: LFM3-1.7B has 1.73B parameters, a 256K-token context window and support for 16 languages, and the company says it fits in about 1 GB of memory at 4-bit precision.
4. Vision model: LFM3-VL-2B takes video input for the first time, and the company says it averages 71.4 across 62 vision benchmarks, close to the 71.7 of Gemma 4 12B, a model more than six times its size.
5. Audio and omni: LFM3-Audio-2B supports transcription, synthesis and translation across 18 languages; LFM3-Omni-2B, the first omni model, can answer spoken questions about images and video.
6. Decision models: d1-2B and d1-Omni-2B answer all questions in a single forward pass with calibrated probabilities, and d1-2B has a third fewer parameters than the previous d1-3B.
7. License and access: the models ship as open weights under the LFM Open License v1.0, free for commercial use below $10 million in annual revenue, with a separate commercial agreement required at or above that threshold.