Docs
OfficialJev Practices & Patterns
Images and Assets (Dreamer-Ready)
The experimental image-input reference: PNG or JPEG files in a top-level assets map, referenced by string anywhere in state, instructions, or criteria, behind the TypeSafe-Danger-Zone-Images header.
- Format
- Docs
- Reading time
- 2 min
- Published by
- TypeSafe AI
- Last verified
- Oct 2026
Highlights
- Images ride in a top-level assets map keyed by reference string and are placed in state, instructions, or criteria; the API swaps each reference for the image.
- PNG or JPEG still images only, 256 to 1568 pixels per side, up to 16 MiB per file, 32 MiB per request body shared after processing.
- Both jev-1.13.0 and jev-1.14.0 accept images; every request with an image or Box question needs TypeSafe-Danger-Zone-Images: true.
- The page marks itself dreamer-ready: request format, answers, and limits can change between releases, and accuracy changes as models update.
Watch out
Experimental and explicitly not production-ready: the docs say to use it for prototypes, warn that all models perform worse on images than on text, and recommend decomposing questions into the simplest possible form. Limits and request shapes can change between releases.
Questions about this entry
Short answers drawn from the linked project or page.
Each image is an entry in the top-level assets map, keyed by a reference string, and that string is placed anywhere the JSON interface allows: in the state, in instructions, or in criteria. The API replaces every reference with its image before the model reads it.
Both current models: jev-1.13.0 and jev-1.14.0. The header TypeSafe-Danger-Zone-Images: true is required on every request that contains an image or a Box question, and the page notes the support is experimental.
PNG or JPEG stills, 256 to 1568 pixels per side, up to 16 MiB per image file and 32 MiB per request body shared by every distinct image after processing. Animated or damaged images, unreferenced assets, and invalid base64 are rejected.
The page warns that all models, Jev included, perform worse on multimodality over pure text, and that breaking questions into the simplest possible decompositions helps a lot. Treat image questions as prototypes until TypeSafe removes the danger-zone gate.
No. Images bill as input tokens at the normal per-token price; there is no separate image charge. The release post and MarkTechPost's summary both confirm the standard $0.042 per million input tokens applies.
Yes. The composable find-and-replace works in state, instructions, or criteria, so an option's description or a Noul's true/false criteria can reference an image the same way state fields can. Each use counts toward the limits again.
The request fails before any model runs: the API returns HTTP 400 telling the caller images are experimental. The same 400 covers a Box question sent to a model that cannot answer it, such as jev-1.13.0 or jev-latest.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
X: First among models under 10B on the Decision Index v0.2.1. Liquid AI's d1-3B is built from LFM2.5-VL-3B and…
X: The classic hotdog classifier as a real-time vision decision. Fayaz Ahmed rebuilt the not hotdog app on…
X: Liquid d1 with vision
More Jev resources like this
The box question type: one bounding box per question, returned as four edges that are fractions of the image, plus confidence, with no criteria field and no not-found value.
Jev Practices & PatternsDocs#official#docs#vision
Official
Documents a failure mode in letter-bound remote sensing classification: 97.71% OA on UCM in vocabulary but 7.35%–96.40% cross-dataset accuracy and inverted confidence (AUROC 0.38). RSAJEV replaces the letter readout with training-free prototype retrieval, reaching 89.43% on the UCM cross-dataset gate.
Jev Practices & PatternsArticle#community#research#arxiv
Community
An evaluation of Jev and two open-weight typed decision models when option names and definitions conflict. With yes/no names, decision-flip rates are up to 70.4 percentage points higher than the 0/1 control, and mean AUC drops from 93.8% to a below-chance 23.2%.
Jev Practices & PatternsArticle#community#research#arxiv
Community
The Jev Day release post: Jev 1.14.0 under jev-preview plus three danger-zone features — composable images on both models, the box question type on 1.14 only, and opt-in 262k context.
Jev Guides & ArticlesArticle#official#release#vision
Official
0xRinne's MIT-licensed field report on using TypeSafe Jev 1.14's Box type and image input to spot AI-image hallucinations, across seven pipeline versions v1-v7b, with a live Vercel demo and raw score JSON.
Jev Tools & CLIs#community#research#hallucination-detection
Community
RINNECODER/jev-box-studyJavaScript
RINNECODER/jev-box-studyJavaScriptA study of option-order sensitivity in System One models and a training-free fix called Runoff. It scores each option alone to keep a few candidates, then compares them with each candidate in every position, removing order dependence in two forward passes.
Jev Practices & PatternsArticle#community#research#arxiv
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.