最新模型
Liquid AI Releases d1-ultra, a Multimodal Decision Model That Returns Option Probabilities in a Single Forward Pass
On October 15, Liquid AI released d1-ultra, which accepts text and images and directly returns probabilities for classification or scoring results. The release emphasizes challenging tasks and probability calibration, but performance in Chinese-language settings and real-world latency still require validation by deployers.
Liquid AI released d1-ultra on October 15, applying a multimodal model to decision tasks with fixed options. The model accepts text, images, and questions, then returns probabilities in a single forward pass rather than generating an answer token by token. The company says the new version is based on a larger open-weight model, with custom modifications and training. Its release channel is a hosted API, so that does not mean d1-ultra itself has open weights. Launch announcement
The interface retains three question types: Noul returns probabilities for yes-or-no questions, Choice provides a probability distribution over candidate labels, and Score returns a result according to an ordered scoring rubric. The official documentation lists request routing, content classification, agent tool-call review, and image inspection as use cases. Engineering teams could evaluate connecting fixed-answer workflows to this decision interface to reduce the steps involved in generating text and parsing outputs; whether this lowers end-to-end latency still needs to be measured. Model documentation
There is also a pricing detail that is easy to overlook. The announced price is $0.042 per million input tokens, with no output-token charge, but each question separately counts the text and all images. So when one image is used to answer multiple questions, costs still increase with the number of questions. Estimates should therefore be based on the complete workflow. Pricing details
Evaluation should consider not only how often the model is correct, but also whether its probabilities are trustworthy. Artificial Analysis has added d1-ultra to its Decision Index, assigning equal weight to four evaluations: knowledge and hallucinations, multiple-choice reasoning, tool calling, and safety classification. It also reports calibration, latency, and cost separately. Calibration plots compare model confidence with actual accuracy and can help determine whether probabilities are suitable for setting automated-processing thresholds; an aggregate score still cannot represent performance on every business dataset. Evaluation methodology and leaderboard
For Chinese-language deployments, the next things to examine are classification and calibration performance with Traditional Chinese, domain terminology, and real-world image quality. Based on the interface’s design, engineering teams can start by comparing error rates on existing labeled data, then validate thresholds for routing low-confidence cases to human review and measure the total cost and tail latency of multi-question requests. These are pre-deployment validation steps; available sources do not yet establish an advantage in Chinese-language settings.