TypeSafe AI Releases Jev 1.14 With Experimental Image, Box and Long-Context Features

Jev is not a text generator. It returns typed decisions: developers send a state together with typed questions — Choice, Score or Noul — and the model answers each with a typed value, probabilities and a confidence score. TypeSafe trains it with Reinforcement Learning for Calibrated Decisions, or RLCD. According to the release post, version jev-1.14.0 brings small, subtle improvements to almost everything, with the largest gain expected in multilingual support. TypeSafe published no public scores for that claim.

All three new features sit behind request headers, and TypeSafe states that none of them is ready for production. Images are composable: each image is listed once in a top-level assets map under a reference string, and that reference can be placed anywhere in state, instructions or criteria; the API swaps the reference for the image before the model reads it. Requests need the header TypeSafe-Danger-Zone-Images: true. The documentation allows PNG or JPEG still images only, up to 16 MiB per file, 256 to 1568 pixels per side and 32 MiB per request body. Images bill as input tokens at the normal price. TypeSafe says jev-1.13.0 was multimodal all along and that both 1.13 and 1.14 accept images. It also warns that all models, Jev included, perform worse on images than on pure text, and recommends the simplest possible questions.

Box is a new question type that returns one bounding box per question, with top, bottom, left and right edges expressed as fractions from 0 to 1, plus a confidence score. It requires jev-1.14.0 or jev-preview and the images header. Box always returns a box — a question about an absent object still yields four edges — so TypeSafe suggests pairing it with a Noul question that checks whether the target exists. In a test on 36 painting crops that each cut off part of the target, four per-edge Noul questions caught 30 cuts, a single "is all of it inside" question caught 26, and drawing the box as an outline and asking about it caught 20.

Context behavior changes through an opt-in header rather than a new default. Both models keep 64k tokens per request and 32k per question; sending TypeSafe-Danger-Zone-Long-Context: true raises both limits to 262k tokens. TypeSafe states that Jev always supported longer context internally but capped it on purpose because the model, like all models, is less reliable on long inputs. Its developers still recommend short contexts.

The release post, reported by MarkTechPost, includes internal intelligence charts that compare Jev only against open-weight models and contain no numeric score table. TypeSafe says the measurement is likely biased toward its own models and tells developers to favor their own evaluations. The post also includes a comparison table placing Jev 1.14 against Jev 1.13 and Gemini 3.5 Flash-Lite, which is listed with a 1,048,576-token input context and prompted JSON box_2d output, against Jev's native box type.

Pricing is unchanged at $42 per billion input tokens, or $0.042 per million; output tokens are free. The model runs only on TypeSafe's hosted API at POST /v1/systemone, with no open weights; parameter counts, hardware and quantized variants are not disclosed. Under the rollout plan, jev-preview points to jev-1.14.0 now and jev-latest moves to it in a week. jev-1.13.0 remains available with 60 days notice before deprecation, and jev-1.14.0 will get 30 days notice before its own.

Editor's Summary

TypeSafe AI's Jev 1.14 is live under the jev-preview alias with three experimental capabilities — image input, a native bounding-box question type and an opt-in 262k-token context — none of which the company says is production-ready. Input tokens remain priced at $0.042 per million with free output, and jev-latest is scheduled to switch to the new version in a week while jev-1.13.0 stays available ahead of deprecation.