Liquid AI launches d1-ultra, a decision model that answers without generating text

Liquid AI released d1-ultra on October 15, 2026, a multimodal decision model aimed at hard, messy real-world checks such as spotting a defect on a circuit board against its spec or comparing a damage claim with a customer's photo.

Decision models work differently from chat models. Instead of writing an answer token by token, d1-ultra reads text, images or both along with one or more questions in a single pass, and returns a probability for each possible answer. It handles three question types: yes or no, picking one label from a list, and placing something on a scale. Liquid says it built d1-ultra on a larger open-weight base model than its earlier d1, with heavy customisation, and trained it specifically to give well-calibrated probabilities, since it found most decision models are overconfident.

In Liquid's comparison across 112 evaluations, d1-ultra beat its own d1 and the rival Jev 1.14 on every task at the same price, and the company says it leads all other decision models on the Artificial Analysis Decision Index by at least 8 points. It costs $0.042 per million input tokens, with no output tokens billed. Images count as input tokens, so a 1024 by 1024 image costs 1,536 tokens. It is on Liquid's API now, with Vercel and OpenRouter to follow.

Why it matters

For classification, routing and inspection tasks, a decision model can be far cheaper and faster than prompting a full chat LLM and parsing its reply.

Source: Liquid AI