Decision models stop at right or wrong: d1-ultra trains calibration too, beating Jev 1.14 on accuracy and calibration across 112 evaluations

Decision models now have to return a trustworthy probability alongside the answer, not just the right label. Most decision models work well on easy and common tasks but are overconfident in their responses, so their probabilities are unreliable. Against the previous d1 and Jev 1.14 across 112 unique evaluations, d1-ultra gives higher accuracy and better-calibrated answers at the same price, and leads other decision models on the Artificial Analysis Decision Index by at least 8 points.

The approach builds on a larger open-weight model with heavy Liquid-specific customization and explicit training for probability output. d1-ultra keeps the three question types of d1: noul returns a yes/no probability from 0 to 1, Choice returns one label from many with a probability per label, and Score returns a position on a scale. The model reads text, images, or both at once, for example a production-line circuit board photo plus its spec asked whether a defect is present; images count as input tokens at 1.5 tokens per 32×32-pixel patch, so a 1024×1024 image costs 1,536 tokens.

Intended for developers and engineers who need probabilistic answers in real-world settings, called through the Liquid API at $0.042 per million input tokens, with images priced the same as text. The model will soon be available through Vercel and OpenRouter.

The evaluation covers 112 tasks, but the material does not give their distribution or difficulty mix; the calibration improvement is presented as an overall comparison rather than per task.