Liquid AI's d1-ultra decision model takes a very strong place on the pareto front

Liquid AI just introduced a very strong, very fast and quite reasonably priced decision model: d1-ultra. Liquid AI has released d1-ultra, a multimodal decision model that reads text, images, or both in a single forward pass and returns calibrated probabilities for yes/no, multi-label choice, and scale questions, without generating tokens. Built on a larger open-weight base than its predecessor d1 and tuned with Liquid-specific customizations, it targets harder, messier real-world tasks such as checking a circuit board against its spec or validating a damage claim against a customer photo. The company reports gains over d1 and Jev 1.14 across 112 evaluations, with particular improvement on knowledge-related tasks, and says the model is better calibrated, since decision models tend to be overconfident. It is available now through the Liquid AI API at $0.042 per million input tokens, with Vercel and OpenRouter listings planned, and the release states that d1-ultra outperforms other decision models on the Artificial Analysis Decision Index by at least 8 points. https://www.liquid.ai/blog/d1-ultra https://x.com/liquidai/status/2110777964489285875 https://artificialanalysis.ai/evaluation