Jev 1.14 Now Accepts Images

What exactly arrived in Jev 1.14

I would read this update as follows: Jev 1.14 can now accept static images and use them in questions that drive decisions. By early November 2026, the capability had been available for around two weeks, so this is no longer breaking news but a useful technical review. In TypeSafe AI's announcement and Cloudflare's documentation, the feature is explicitly described as experimental.

PNG and JPEG are supported as input formats. GIF, WebP, HEIC, video, and animated PNG files are not accepted directly. According to the available documentation, a single image is limited to 16 MiB, while the full request body may not exceed 32 MiB.

An image can be sent as a separate multipart request part or embedded as base64. The asset is then linked to a field in state referenced by questions. Activation requires the TypeSafe-Danger-Zone-Images: true header, whose name rather honestly reflects the feature's maturity level.

The main functional addition in version 1.14 is the Box primitive. It is intended to find a specified object and return its region in an image. Visual data can also be used in Choice, Score, and other decision-making questions.

There is an important boundary here: integration examples involving a separate vision model do not prove that Jev independently performs the entire visual analysis. In those setups, the vision model first produces structured evidence, and Jev then makes a decision based on it.

What this changes for integrations

The practical impact is real, but reliable production use is still some distance away. Images can now be included directly in task state, reducing the amount of intermediary code between visual analysis and evaluation logic.

I would first test more than the demo examples: Box stability, behavior near size limits, and the reproducibility of Choice and Score results. A separate feature flag, beta status, and an explicit warning against production use all indicate that neither the API contract nor response quality should be considered settled.

There is no confirmed independent official leaderboard for these visual capabilities in the available materials. This is therefore not a ready-made universal vision engine, but an early extension of a decision API that becomes much more interesting at the intersection of images and structured decision-making.

We previously covered Pydantic Monty, a secure Python interpreter designed for LLM-driven code execution. It provides useful context for how Pydantic is expanding the technical capabilities around AI applications, including multimodal model inputs.