TypeSafe has released Jev 1.14.0, an update to its Jev model family, alongside a set of experimental capabilities the company is calling “danger zone” features. These include image support, a new structured output type called box, and the option to use much longer context windows than before. Rather than bundling everything into one polished, production-ready release, TypeSafe is explicitly splitting its roadmap into reliable tools for “prod builders” and riskier, more flexible tools for developers experimenting with new ideas, a distinction that matters because it signals how the company plans to balance safety with innovation going forward.

What is actually new in Jev 1.14.0

The headline release is Jev 1.14.0 itself, a successor to Jev 1.13.0. TypeSafe is also shipping jev-preview, a separate version meant to deliver small, broad improvements rather than flashy new capabilities, with the company specifically calling out multilingual support as the area of biggest expected gains. Importantly, jev-latest will not switch to 1.14.0 immediately; TypeSafe says that transition happens after a one-week delay, giving developers a short buffer to test before the default model changes under them. The older 1.13.0 version will remain available, with TypeSafe promising 60 days of warning before deprecating 1.13.0 and 30 days for 1.14.0, backed by a deprecation policy to be published later.

Why images were already there, and what changed

One of the more surprising disclosures in the announcement is that Jev 1.13.0, the version developers have already been using, was multimodal all along. TypeSafe is now formally enabling and documenting that image support rather than introducing it fresh. Images can reportedly be inserted anywhere in the JSON interface that Jev accepts, including in the state, instructions, or criteria fields, using what the company describes as a find-and-replace mechanism in assets. This is a flexible, composable approach compared to systems that restrict images to a single fixed input slot. However, TypeSafe is upfront that multimodal inputs make the model less reliable than pure text, and it recommends breaking complex visual questions into simpler sub-questions to get better results.

A new type called box, and why it matters for structured outputs

Jev 1.14.0 introduces a new output type called box, which returns a single bounding box with no type errors allowed. TypeSafe frames this as a notable shift: previous types in its system were essentially built for “smart if-statements,” meaning structured decisions or classifications. Box is described as the first type aimed at something beyond simple decision-making, hinting at broader ambitions for what the company calls “System One Models” to handle spatial or visual reasoning tasks. Unlike the image support, box is only available on jev-1.14.0, not on the earlier 1.13.0 release, which the company attributes to recent internal progress on the feature.

Longer context, opt-in telemetry, and the reliability tradeoff

TypeSafe confirms that Jev already supports a longer context length internally than what has been enabled for production use. The company says it intentionally capped context length in the past because longer contexts tend to reduce model reliability, and the limit nudged developers toward more effective usage patterns. With this release, that cap becomes optional rather than mandatory: developers can now choose to use longer context windows, with TypeSafe comparing its stance to not being “your parents” even while still recommending shorter contexts for most use cases.

Separately, TypeSafe is adding an opt-in setting that lets developers share their queries to help the company identify weaknesses in the model, sometimes referred to in the industry as model “jaggedness,” meaning inconsistent or unreliable behavior across similar tasks. The company states it will not train on these shared queries, framing the feature purely as a debugging aid rather than a data collection mechanism.

What this means in practice for developers

For teams building production systems, the practical takeaway is that nothing is forced to change immediately. The default jev-latest alias will update to 1.14.0 after a delay, and the older 1.13.0 model sticks around with a defined deprecation runway, which should limit sudden breakage. For developers who want to experiment, the new image handling, the box type, and extended context are accessible through what TypeSafe calls danger zone flags in the API headers, explicitly marked as not vetted for production reliability. Documentation for the image features and context length settings is available on TypeSafe’s docs site, linked in the original announcement, for developers who want to try these capabilities directly.

TypeSafe also reiterates a longstanding stance against relying on public benchmarks, saying it instead uses internal measures aimed at capturing the kind of intelligence that matters most to developers, though specific benchmark results were not detailed in the excerpt available.

Frequently asked questions

Do I need to update my code to keep using Jev? No. jev-1.13.0 remains available, and jev-latest will not point to 1.14.0 until a week after this announcement, giving developers time to test.

Are the new danger zone features safe for production? TypeSafe explicitly says no, describing them as not yet battle-tested to the company’s reliability standards, intended instead for experimentation and creative use cases.

What is the difference between jev-1.14.0 and jev-preview? jev-1.14.0 is the main release with new capabilities like the box type, while jev-preview focuses on broad, incremental improvements, with multilingual support expected to see the largest gains.

Source: We’re releasing Jev 1.14.0 along with experimental features for dreamers