TypeSafe releases Jev 1.14.0 with opt-in image, box and long-context flags
TypeSafe has announced Jev 1.14.0, released under jev-preview, together with a set of experimental features aimed at what it calls 'dreamers'. The post frames the split as a balance between reliability and creativity. Production builders get models that meet the company's quality bar; experimenters get features shared under 'danger zone' flags in the API headers. TypeSafe says these features will stay for future models and are, as far as it can tell, better than what exists elsewhere, but it thinks there is still way too much jaggedness to vouch for them in production. It describes them as a continuation of over 2 years of stealth work.
Images come first. TypeSafe says the current model, jev-1.13.0, was multimodal all along, and multimodality is now enabled for both current and future Jev models. Images are composable through a find-and-replace in assets and can go anywhere the JSON interface supports: in the state, instructions or even criteria. They are switched on with the header TypeSafe-Danger-Zone-Images: true. The company warns that all models, Jev included, perform worse on multimodality than on pure text, and that breaking questions into the simplest possible decompositions helps a lot.
Second is a new type called box, which returns a single bounding box. It is the first Jev type not purely meant for smart-if statements, and the company calls it a preview of exploration beyond simple decisions into what else System One Models can support. Box is enabled only for jev-1.14.0.
Third is context length. Jev supports a much longer context than is enabled in production; TypeSafe says it limited it intentionally because Jev, like all models, is less reliable on long contexts and the limit encourages better usage patterns. It still strongly recommends limiting context, but users can now lift the limit with TypeSafe-Danger-Zone-Long-Context: true.
There is also an opt-in that lets users help find model jaggedness with their queries. TypeSafe says it will still not train on queries, but the opt-in makes it easier to find issues and debug use cases.
jev-1.14.0 itself is described as not flashy: small, subtle improvements to everything, with the largest gain expected in multi-lingual support. jev-latest will be updated to jev-1.14.0 in a week, and jev-1.13.0 will stay available for a while. TypeSafe intends to give 60 days warning before jev-1.13.0 is formally deprecated and 30 days for jev-1.14.0, and says it will release a deprecation policy before that.
The post also shares an internal 'intelligence' measurement, noting the company dislikes public benchmarks. Its caveats: the measurement is a best guess and will keep evolving, readers should favor their own evals, only open-weight models are included because the company guards its validation set closely, and the result is likely biased towards its own models. TypeSafe suggests focusing on the relative ranking of the open-source models. Looking ahead, it says most releases should be 'beautifully boring': small improvements, better stability and features for real deployment.
Key facts
- Jev 1.14.0 ships under jev-preview with small improvements across the board; the largest gain is expected in multi-lingual support.
- Images, a new 'box' type (returns a single bounding box) and longer context are opt-in via 'danger zone' API headers, which TypeSafe says it does not vouch for in production.
- The current jev-1.13.0 was multimodal all along; the box type is enabled only for jev-1.14.0.
- jev-latest moves to jev-1.14.0 in a week; TypeSafe intends 60 days warning before deprecating jev-1.13.0 and 30 days for jev-1.14.0.
- An internal measurement against open-weight models is shared with caveats: it is likely biased towards TypeSafe's own models.
Why it matters
The release shows a vendor deliberately separating a stable production line from experimental capabilities. Instead of holding back images, longer context and a new output type until they meet its reliability bar, TypeSafe exposes them behind explicit header flags. It also says the production model was multimodal from the start, so image input arrives for jev-1.13.0 as well as for jev-1.14.0. The box type is notable because it is the first Jev type not purely meant for smart-if statements, a step beyond simple decisions.
Who it affects
Developers building on the Jev API are the audience. Teams running production workloads get a small-improvement model and a promised upgrade path; experimenters and builders of prototypes get the danger zone features. Anyone pinned to jev-latest will see it move to jev-1.14.0 in a week, and those who want to stay put can use jev-1.13.0, which TypeSafe says will be available for a while.
How to use it
Each experimental feature is a request header. Image support is TypeSafe-Danger-Zone-Images: true, with images placed through a find-and-replace in assets anywhere the JSON interface allows (state, instructions or criteria). Long context is TypeSafe-Danger-Zone-Long-Context: true. The box type is available only on jev-1.14.0. The post points to docs.typesafe.ai/go/images, docs.typesafe.ai/go/box and docs.typesafe.ai/models#context-length for details. To pin a version, jev-1.13.0 stays available. TypeSafe also suggests breaking multimodal questions into the simplest possible decompositions.
How solid is it
This is a first-party announcement on the company's own site, so every claim is TypeSafe's own. The post gives no actual benchmark or internal measurement numbers, only the caveats around the measurement. It states the measurement is a best guess, likely biased towards TypeSafe's models, and tells readers to favor their own evals. It also gives no specific context-length figures, and no calendar date for the release. The company itself says the danger zone features are not ready for production use.
Risks and caveats
TypeSafe warns of 'way too much jaggedness' in the experimental features and says all models perform worse on multimodality than on pure text. Longer context is less reliable, which is why the company limited it in the first place and still strongly recommends limiting it. The deprecation window for jev-1.14.0 is intended to be 30 days, half of the 60 days promised for jev-1.13.0, and the deprecation policy has not yet been published. The query opt-in shares queries with the company for debugging, though it says it will still not train on them.
“we think there is still way too much jaggedness for us to vouch for these for production use cases!”
— TypeSafe, Jev 1.14.0 release post