Typesafe AI launches Jev 1.14 with image support, box type and expanded context limits

TypeSafe AI has launched Jev 1.14, a new version of its System One decision model, now available under the jev-preview alias. The update introduces three experimental features: image support, a box type, and extended context limits. These features are labeled as 'danger zone' and are not yet production-ready. Jev, a typed decision model, returns structured outputs instead of generated text, allowing developers to build applications for tasks like inspection and localization. The new version enhances multilingual support and allows images to be embedded via references in the state, instructions, or criteria. Images are processed by the API before the model reads them. The box type returns bounding box coordinates as fractions and confidence scores, requiring the images header for activation. TypeSafe recommends pairing box queries with Noul questions to check for object presence. Jev 1.14 also increases context limits to 262k tokens with the 'long-context' header, though TypeSafe advises against using long inputs due to potential unreliability. The company warns that models perform worse on images than text and recommends simple questions. Pricing remains unchanged at $0.042 per million input tokens.