TypeSafe AI's Jev 1.14.0 adds image input, a bounding-box type and longer context

Jev 1.14.0 arrives with opt-in experiments

TypeSafe AI released Jev 1.14.0 on October 14, in a blog post that quickly reached Hacker News's front page. The company describes the new version as a round of small, steady refinements rather than a flashy upgrade, with the largest expected gain in multilingual performance. The headline items are actually experimental: image input, a new type called box, and an unlocked longer context window, all gated behind what the company labels "danger zone" flags in its API headers.

The naming reflects a tension the team says it has debated internally: production developers need dependable models, while more experimental users want capabilities the company cannot yet vouch for. TypeSafe AI's answer is to ship those features early — it believes they already beat what else is on the market — while making clear they are still too inconsistent to recommend for production use.

Composable images

The most surprising claim in the announcement is that the previous release, jev-1.13.0, was multimodal internally all along; the capability simply was never exposed. With the new TypeSafe-Danger-Zone-Images header enabled, image support works through a find-and-replace mechanism over assets, so images can be placed anywhere the JSON interface accepts them — in state, in instructions, or even in criteria.

TypeSafe AI adds a caveat: all models, Jev included, perform worse on multimodal tasks than on pure text, and the company suggests breaking questions down into the simplest possible steps to get better results.

The box type

Jev also gains a new type, box, which returns a single bounding box. According to TypeSafe AI, it is the first type not designed purely for the model's smart-if decision flows, and it previews a move beyond simple decisions into what the company calls System One Models. The type is enabled only on jev-1.14.0.

Longer context, unlocked

The company says Jev already supports a much longer context than production exposes, and that the cap was deliberate: models are less reliable over long contexts, and the limit encouraged more effective usage patterns. That constraint is now optional. Developers willing to accept the risk can turn on the extended window with the TypeSafe-Danger-Zone-Long-Context header.

Data sharing opt-in and versioning

A separate opt-in lets users contribute their queries so the team can hunt for model inconsistencies more effectively. TypeSafe AI says it still will not train on those queries, but the data makes it easier to find bugs and debug individual use cases.

On rollout, jev-latest will move to 1.14.0 in a week, while 1.13.0 remains available for now. The company plans to give 60 days of warning before 1.13.0 is formally deprecated and 30 days for 1.14.0, with a full deprecation policy to be published beforehand.

Internal numbers, shared with caveats

TypeSafe AI also published its internal measurements of model intelligence while distancing itself from public benchmarks. The ranking covers only open-weight models — the company says it guards its validation set closely and distrusts the data policies of its model providers — and it acknowledges the measurement is likely biased toward its own model, will keep evolving, and should not be trusted over users' own experiments and evaluations. Readers are directed to focus on the relative ordering of the open-source models instead.

Why it matters

The "danger zone" pattern is a notable experiment in release strategy: rather than holding back unfinished capabilities or pushing them to everyone, TypeSafe AI exposes them behind explicit, honestly labelled opt-in flags. The revelation that 1.13.0 was quietly multimodal all along suggests vendors sometimes ship models with capabilities left dormant. The box type hints at Jev expanding from decision-style outputs toward perception tasks such as object localization. And concrete deprecation timelines — still rare among model providers — give developers a predictable upgrade path, which matters as more teams pin production workloads to specific model versions.