TypeSafe Jev 1.14.0 Introduces Multimodal, Box Detection, and 262K Context Behind Danger Zone

Release Overview

TypeSafe released Jev 1.14.0 on October 14 at 1:14 PM Pacific Time, a version number matching the timestamp. The official blog titled "For the builders AND dreamers" explains the team's internal debate between reliability for production users and creative freedom for experimenters. Their solution: experimental features reside in a "danger zone" enabled via specific API request headers, retained and improved but not guaranteed production-ready.

Multimodal Support

Jev 1.13.0 already possessed multimodal capabilities but kept them private due to reliability concerns outside text. Version 1.14.0 exposes image input via the TypeSafe-Danger-Zone-Images: true header. Images are referenced in the assets field and can be used in state, instructions, and criteria. The documentation warns that all models are less reliable on multimodal tasks than on pure text; developers should keep image tasks simple and avoid complex visual reasoning.

Box Output Type: Localization, Not Judgment

Jev introduces its first non-decision output type: box. Unlike previous outputs (boolean, multiple choice) designed for smart-if, box returns the bounding box of a target in an image. Coordinates are proportions of image width/height from the top-left corner, e.g., left: 0.44, right: 0.84, top: 0.25, bottom: 0.82. The format guarantees no type errors and returns exactly one box per call. Available only on jev-1.14.0.

A notable design choice: the model always returns a box even if the target is absent. The docs recommend pairing with a Noul type to first verify presence, then localize. Practical tips include adding padding when confidence is low, cropping and upscaling for a second pass, or querying each edge individually to check overflow. This enables automation workflows for quality inspection, compositing, and video editing where precise coordinates replace vague descriptions.

262K Token Context: Available but Discouraged

Adding the TypeSafe-Danger-Zone-Long-Context: true header raises both per-request and per-question context limits to 262K tokens (defaults: 64K per request, 32K per question). TypeSafe states the model always supported longer context but deliberately limited it because long context reduces reliability. They advocate careful state management over "mindless context stuffing." The blog quips, "We aren't your parents though," implying developers may enable it at their own risk. Longer context is slower and less accurate; use only when genuinely needed.

Additional Details

Opt-in data sharing: users can allow queries to help TypeSafe detect model "jaggedness"; data is not used for training but for debugging and issue localization.

Version aliases: jev-preview now points to jev-1.14.0; jev-latest switches in one week. jev-1.13.0 retained with 60-day deprecation notice; 1.14.0 will have 30-day notice. Pricing unchanged: $42 per billion input tokens, output free.

Multilingual support is the largest improvement in 1.14.0; other areas see marginal "everywhere a little better" gains.

TypeSafe publishes internal benchmarks with a prominent warning: "AI benchmarks. Please don't trust them." Only open-source models are compared due to closed-source data policy concerns; evaluations may favor Jev. Developers are urged to run their own tests.

Commentary

The author views TypeSafe's approach as restrained: experimental features are opt-in via headers, clearly marked non-production, avoiding default exposure that could cause production incidents. The box type offers the most imaginative potential, turning "seeing" into "locating" and closing the last mile for many automation pipelines. However, the repeated "danger zone" warnings mean production systems relying on Jev for decisions should not rush to upgrade. Test with jev-preview first, wait for jev-latest to roll over. Jev's $7.5B valuation means adoption will determine whether this update delivers on its promise.