TypeSafe teaches Jev to see: images now work in state, instructions and criteria, alongside Jev 1.14, a box type and 262k context
The standout is image input. On 14 October 2026 TypeSafe published “Jev: For the builders AND dreamers”, and its most notable change is that images are now a first-class, composable part of a System One request. An image can go anywhere text can: in the state, inside a question’s instructions, or as the criteria a question is judged against (for example, “is the dog in photo the same dog as in this poster?”). Every question type accepts images, so a single request can mix photos with typed questions and get calibrated probabilities back. Two points of accuracy: per TypeSafe, jev-1.13.0 already supported images (“multimodal all along”), so this is a documented, opt-in capability on the current model, not a new model; and it is a danger-zone feature (TypeSafe-Danger-Zone-Images: true) that TypeSafe says is not yet ready for production, with request format and limits that can change. TypeSafe also warns that all models do worse on images than on text.
Alongside it:
- Model release: small.
jev-1.14.0, under thejev-previewalias, with “subtle improvements to everything” and the biggest gain expected in multilingual.jev-lateststays on 1.13.0 and moves to 1.14.0 about a week later. Price is unchanged ($0.042/M input, output free). TypeSafe promises 60 days’ notice before deprecating 1.13.0 and 30 days for 1.14.0, with a policy to follow.
The images, box and long-context features all sit behind opt-in “danger zone” headers that TypeSafe says are not production-ready.
boxquestions: return one bounding box with a confidence, on Jev 1.14 only.- Long context (
TypeSafe-Danger-Zone-Long-Context: true): 262k tokens per request and per question, up from 64k / 32k, at the same per-token price; slower and less accurate. - An organisation-level opt-in lets TypeSafe read inputs and outputs to debug jaggedness; it says it still does not train on them.
Claim, not evidence. TypeSafe’s own chart (private set, own model, stamped “please don’t trust them”) puts Jev 1.14.0 at 60.0 Intelligence vs 58.9 for 1.13.0, and calibration error 1.6 vs 1.8; the next model is at 46.7. These are an UnverifiedClaim and the release is not on JevBench, Decision Index 0.3 or the Artificial Analysis index yet. We will update when it is.
| Model | Intelligence (higher better) | Calibration error (lower better) |
|---|---|---|
| jev 1.14.0 | 60.0 | 1.6 |
| jev 1.13.0 | 58.9 | 1.8 |
| kev-27b | 46.7 | 17.8 |
| pplx-decider-v1.1-27b | 46.0 | 35.0 |
| clef | 44.3 | 11.2 |
| clef-flash | 38.8 | 19.1 |
| kev-4b | 36.9 | 10.3 |
| d1-3b | 33.8 | 19.9 |
Chart (UnverifiedClaim): private set, closed Jev included despite “only open weight models”, “likely biased towards our models”, no confidence interval.
Catalog page: Jev.