TLDRocket
7 October 2026
The day’s biggest thread is AI getting faster and closer to where it’s needed: Liquid AI just released open-weight multimodal decision models designed for the edge, led by d1-3B. On seven public datasets it posted an average score of 82.9, then—more importantly for real-world deployments—answered a single question in 16 ms on an NVIDIA Jetson AGX Thor. The trick is that the model makes a decision in one forward pass, not a chat detour: d1-3B handles text-plus-image inputs, while the experimental d1-omni-600M extends to text+image or text+audio. It’s now on Hugging Face with usage code, the sort of release that turns “can it run?” into “how do we ship it?”
Read the full briefing →