TypeSafe AI has launched Jev 1.14, updating its System One decision model with experimental support for images, a new box type, and a 262k context window. Available now under the jev-preview alias, the updated typed decision model expands into vision tasks such as visual inspection and localization through experimental danger zone features.
The Jev 1.14 update brings small performance improvements across the board, with the largest gains reportedly in multilingual support. The new image feature allows developers to include PNG or JPEG images up to 16 MiB via a top-level assets map, billed at standard input token rates. Additionally, the new Box question type returns a single bounding box with normalized edge fractions and a confidence score, though TypeSafe warns that the experimental features are not yet ready for production use.
Developers can also opt into an expanded context window of 262k tokens per request by using a special header, up from the default 64k limit. Pricing remains unchanged at $0.042 per million input tokens with free output tokens, while the model runs exclusively through TypeSafe's hosted API.
This release matters because it pushes typed decision AI models further into multimodal and long-context enterprise workflows while highlighting ongoing industry caution around experimental features.