TypeSafe AI ships Jev 1.14.0 with experimental image, box type, and longer context
Original: We’re releasing Jev 1.14.0 along with experimental features for dreamers
Why This Matters
Structured-output AI models adding multimodality and spatial types signals expanding use-case scope.
TypeSafe AI released Jev 1.14.0 on Oct 14, 2026, adding experimental image support, a bounding box output type, and extended context length—all gated behind opt-in API 'danger zone' headers for non-production use.
TypeSafe AI's Jev 1.14.0 drops with three experimental features the company is explicitly calling 'danger zone' additions—not production-ready, but stable enough to share via opt-in API headers. First up: image support. Jev 1.13.0 was apparently multimodal all along, and TypeSafe is now unlocking it. Images slot into the existing JSON interface anywhere assets appear—state, instructions, or criteria—via a find-and-replace pattern. Activate with `TypeSafe-Danger-Zone-Images: true`. Second: a new `box` type that returns a single bounding box with no type errors. It's the first Jev type that goes beyond binary decisions, hinting at broader 'System One Model' ambitions. Third: extended context length, previously capped intentionally because longer contexts hurt reliability. The cap is now optional with `TypeSafe-Danger-Zone-Long-Context: true`. The company is blunt: they still recommend against long contexts in prod. The 1.14.0 model itself (released under `jev-preview`) is described as small, incremental, with the biggest gains in multilingual support. TypeSafe also added an opt-in query-sharing mode for community debugging—no training on queries, but easier issue detection. The company frames this dual-track approach as a philosophy split between 'builders' who need reliability and 'dreamers' who want to push limits.