Liquid AI Launches LFM3 Multimodal Models for On-Device Intelligence

Liquid AI introduced LFM3, a new generation of Liquid Foundation Models designed as compact starting points for specialized multimodal intelligence. The release brings text, vision, audio, omni, and decision models onto a shared backbone at about two billion parameters, allowing developers to select a modality-specific model, fine-tune it on a single GPU, and deploy it on devices, private infrastructure, or the cloud. LFM3-1.7B-Base is the foundation for six 2B-class models: LFM3-1.7B for text, LFM3-VL-2B for images and video, LFM3-Audio-2B for speech and audio, LFM3-Omni-2B for text, images, video, and audio with text and speech output, d1-2B for vision-only structured decisions, and d1-omni-2B for multimodal decision tasks. The company said the base model uses 22 trillion training tokens, including 3.5 trillion mid-training tokens focused on agentic traces and tool calls, extending the context window to 256K tokens. The architecture uses a tri-hybrid design with gated short convolutions, sliding-window attention, and global attention, along with head-wise gated attention, partial rotary position embeddings, and sandwich normalization. Liquid AI said the changes improve performance per parameter, quantization robustness, and long-context speed. LFM3-1.7B is claimed to outperform LFM2.5-2.6B while using 35% fewer parameters, and it remains about 1 GB in memory at 4-bit precision for text inference. It also achieved strong results on WebArena-Verified and long agent tasks compared with larger models. The vision model supports video input and step-by-step reasoning, while the audio model supports speech recognition, synthesis, translation across 18 languages, and audio-to-text or audio-to-speech responses. The omni model integrates perception and generation in one on-device system, and the d1 models are designed for calibrated, structured decisions over support tickets, invoices, images, screens, audio clips, and other inputs. Liquid AI said the models are open-weight and available on Hugging Face, with support in Transformers and tooling for single-GPU fine-tuning. The company also demonstrated robotics applications using LFM3-VL-2B, including orchestration, navigation, and manipulation, with action chunks generated in under 50 milliseconds on embedded hardware.