Liquid AI has released d1-ultra, a multimodal decision model designed for difficult, unstructured real-world decisions such as checking circuit-board defects against specifications or comparing damage claims with customer photos. It accepts text, images, or both, along with one or more questions, and processes them in a single forward pass without generating output tokens. The model returns probabilities for yes/no questions, probabilities for each label in a multiple-choice question, or a score on a scale. Liquid AI says d1-ultra is based on a larger open-weight model than its earlier d1 and includes substantial company-specific customizations and training. The company evaluated it against d1 and Jev 1.14 across 112 unique evaluations, emphasizing that decision quality depends on calibrated probabilities as well as accuracy because many models are overconfident. According to Liquid AI, d1-ultra produces better-calibrated answers and higher accuracy than Jev 1.14 across all tasks at the same price, and leads other decision models in the Artificial Analysis Decision Index by at least 8 points. The model is available immediately through the Liquid API at $0.042 per million input tokens, with no charge for output tokens; Vercel and OpenRouter availability is planned. Text and images are billed at the same input rate, with images counted as 1.5 tokens per 32-by-32-pixel patch, so a 1024-by-1024 image counts as 1,536 tokens. Each question is billed as a separate prompt, including its text and images.
