Liquid AI has launched LFM3, a new family of six open-weight models in the 2 billion parameter class, all built on one shared backbone. The family covers text, vision, audio, omni and decision tasks, and is built for on-device AI.
Every model branches from the same LFM3-1.7B-Base, which was pre-trained on 22 trillion tokens. The six models are LFM3-1.7B for text, LFM3-VL-2B for vision and video, LFM3-Audio-2B for audio, LFM3-Omni-2B, the first omni model from the company, and two decision models, d1-2B and d1-omni-2B, which return calibrated answers rather than generated text.
The backbone uses a tri-hybrid design across 24 layers: 13 gated short convolutions, five sliding-window attention layers and six global attention layers. The text model carries a 256,000-token context window and, according to the company, fits in about 1 GB of memory at 4-bit precision. The base and post-trained models are open-weight and available on Hugging Face, the Liquid AI Playground and the Liquid AI API, with the text model shipped under the LFM Open License v1.0.
On company-reported benchmarks, LFM3-1.7B beats its predecessor LFM2.5-2.6B while using 35 percent fewer parameters, scoring 96.66 on IFEval and 71.39 on BFCL V4 tool calling. The vision model is the first LFM vision model to accept video input. Day-one partners include Ambarella, AMD, Arm, Intel, NVIDIA, OpenRouter, Qualcomm and Vercel, while Mercedes-Benz and Shopify partnered on pre-release testing.
Every model can be fine-tuned on a single GPU, which makes the family practical for specialized local agents. Users wanting to run small models locally should also look at Ollama, and compare small-model options through Gemini. One honest caveat from the model card: because of its small size, the text model is not recommended for agentic coding or complex multi-step reasoning.
