Liquid AI introduced six open-weight LFM3 models built on a shared backbone, with multimodal capabilities and options for single-GPU fine-tuning.
Liquid AI introduced LFM3 on October 27, 2026, as a family of six models spanning text, vision, audio, combined multimodal input and structured decision tasks. The company says the models share an LFM3-1.7B-Base foundation and are available as open weights on Hugging Face.
The lineup includes LFM3-1.7B for text, LFM3-VL-2B for text, images and video, LFM3-Audio-2B for speech and text, and LFM3-Omni-2B for combined audio-visual interactions. Two decision models, d1-2B and d1-omni-2B, are designed to return structured answers with calibrated probabilities. Liquid AI says all six can be fine-tuned on a single GPU. The text model supports a 256K-token context window and, at 4-bit precision, fits in about 1 GB.
Liquid AI reports that LFM3-1.7B outperforms its LFM2.5-2.6B predecessor while using 35% fewer parameters. The company also says the model is about 25% faster at an 8K-token context than LFM2-1.2B. Its technical description details a 24-layer architecture combining gated short convolutions, sliding-window attention and global attention. The base model was pretrained on 22 trillion tokens, with a final 3.5 trillion-token mid-training stage emphasizing curated data, agent traces and tool calls.
The family extends beyond text generation. Liquid AI reports that LFM3-VL-2B accepts video, while its audio models support speech recognition, synthesis and tool calling. The company also presents a robotics example built from LFM3-VL-2B: a manipulation model trained across 15 robot embodiments that achieved a reported 96.2% average success rate across four LIBERO suites. These benchmark results and performance comparisons are claims from Liquid AI’s release. The models and fine-tuning notebooks are available through Hugging Face.