Can Jev process images? Native files, OCR, and limits

The hosted model takes text

The models table still lists Input as Text for jev-1.13.0 and jev-1.14.0. Audio and video stay excluded; convert those to text or structured fields first. The state page describes a string, object, or array and does not document image files. The same weights serve every account, and the card says Jev is not fine-tuned or LoRA-adapted with customer data.

On 14 October 2026 TypeSafe documented image files behind TypeSafe-Danger-Zone-Images: true, for jev-1.13.0 and jev-1.14.0. Box questions need jev-1.14.0 or jev-preview. The Jev Day images article covers the header, the assets map, and the size limits. OpenAI’s follow-up on 29 September 2026 says the Decisions API can take text or images as context. Preview access is limited to selected API customers. The workflows below still turn a picture into text, a filename, or a color question before the Jev call, which remains the path when the images header is off.

Color is a separate limit on the jaggedness page. Questions do better with English color names than with hex or RGB triples, and the model cannot reliably judge whether two hex values are close. The page says to convert the numbers in code and pass a name or a bucket. It also says Jev 1.13 is not trained to generate text, so a caption has to come from another model.

Sources:[34] TypeSafe AI[47] TypeSafe AI[155] TypeSafe AI[156] TypeSafe AI[48] TypeSafe AI[109] OpenAI Developers

Read the writing, then ask Jev

keno’s guide starts from the model card. Local OCR reads a screenshot, and Jev classifies the extracted issue and chooses a support queue. He writes that the visual description comes from the extraction step. OCR covers visible writing. It will miss a photo with no text, and a filename or an unverified alt string can be stale or wrong. The sample evidence record in the long post is an example he wrote, and the labeled comparison he proposes, of rules, OCR plus Jev, and a vision model, has no results in the post. He treats an OCR score, a detector score, an embedding similarity, a Jev probability, and Choice confidence as different numbers. A confident misread can become a confident wrong queue.

jev-chat-windows is a working screenshot path for WeChat on Windows 4.x. At commit 033d6ef the app captures the window, runs RapidOCR locally, asks Jev seven questions, has a language model draft three replies, and asks Jev to rank them. Pasting into the input box is manual. The default judgment model is typesafe/jev-1.13 through OpenRouter, or jev-latest on the TypeSafe API. Writing on photos and stickers is dropped when it does not sit on a flat background. The README’s 250 to 800 ms range is one OCR pass in the capture process.

村山毅 Tim reported a separate trial the same week. Latency and cost were about half of Gemini, and accuracy about 80 to 90 percent of Gemini. A follow-up says the accuracy figure is field extraction after OCR. Character recognition on its own is a different measurement. The posts leave out the page count, the Gemini version, and the Jev version, so the pair of percentages stays his report of one unnamed set.

Sources:[49] keno[50] keno[2] TypeSafe AI[51] rezoch340[62] 村山毅 Tim[63] 村山毅 Tim

A mockup that reads filenames

Jeremy Person says a mockup on his own machine uses image filenames and no metadata. He points to Stefan’s video, captioned "when a designer gets access to Jev", as the inspiration. Jeremy’s text never says the files were opened as images. Stefan’s caption leaves the inputs undescribed too. keno’s guide treats a filename as weak evidence for the same reason: it can be stale or simply wrong.

Sources:[52] Jeremy Person[49] keno

One color question per pixel

Anshu’s Jevinci post says the paintings are JavaScript and that no image-generation model is involved. Jev picks a color for each pixel in parallel, and a more confident pick paints a wider brush. The video is marked sped up 4x. Later that day he wrote that the state and the questions took trial and error with Codex, that Jev still answers the pixels in one shot, and that the painting is derived from the grid and the probabilities.

The repository linked from the post now resolves to achimala/jev-paint. Commit ecf9c48 calls jev-latest from the browser. Palette mode asks one Choice per pixel over 16 named colors. Other modes ask for HSL hue, saturation, and lightness, a Noul for each RGB channel, or a silhouette Noul. Grids are 8, 12, 16, 24, or 32 pixels on a side, and the questions are split into batches. The README says HSL and RGB channels are combined as if they were independent, because Jev does not return a joint distribution, and that the renderer is an artistic transform. A loopback Python helper only forwards the request, because the API disallows browser CORS origins. Painting time and the model call are timed separately.

Sources:[55] Anshu[56] Anshu[57] achimala[58] achimala

Jev decides whether the toy should draw

Dragos Dunica built a dinosaur toy for his nephew on a Waveshare ESP32-S3 Touch AMOLED 1.8. Gemini Flash transcribes and carries the conversation. Gemini 3.1 Flash TTS is the voice. FLUX.2 Klein 4B draws. Jev 1.13 decides whether a picture would help and whether the proposed scene is suitable for the child. He calls it a prototype, built with Astra, with the other models reached through OpenRouter.

On Sep 23 he wrote that image generation runs in parallel with the reply. After transcription, Jev decides and calls flux.2-klein-4b. He says the picture is usually finished before the answer is. The slow part, in his account, is transcription and the reply, and he is considering Gemini 3.8 Live instead of a transcription model plus TTS. The thread gives no latency distribution and no count of scenes Jev turned down.

Sources:[53] Dragos Dunica[54] Dragos Dunica

The image-search video’s caption

Shridhar’s Sep 20 post is a 22-second video. The caption calls it fast image search with Jev and stops there, with no mention of pixels, filenames, or descriptions. Under that post, Adam Holter wrote that Jev has no image input, so the images would have to become text descriptions first. Xenova’s quote post makes the same modality point and shows CLIP searching 25,000 images in under 50 ms, with no labeling step, in the browser on WebGPU through Transformers.js. The 50 ms figure is Xenova’s CLIP demo.

Sources:[59] Shridhar[60] Adam Holter[61] Xenova

A logprob bridge and a local image model

LogJev, linked from a reply on keno’s guide, is a separate server. Image and audio parts go to an OpenAI-compatible chat model that returns first-token logprobs, and only when that model supports the input. A provider set up as official Jev forwards the TypeSafe payload. The README says those bridge probabilities are uncalibrated. The accuracy table in the same file is a 144-question text set from the Python implementation.

Ed Plese posted a local model he described as Jev for images. On one 512 by 384 photo he reports 63 ms, and says that is fast enough for 15 fps video. The labels on that photo are Birthday 94%, Golden retriever 78%, Sprinkles 93%, Candles lit 67%, and the dog about to eat the cake at 52%. A follow-up shapes image reports Left at 98% for a blue circle relative to a red square. Both posts give those scores for his local model.

Sources:[64] さよ☆マギカ[65] DumoeDss[66] Ed Plese[67] Ed Plese

Questions about Jev + Images