Up until a couple weeks ago, JEV only worked on text. Now it can look at pictures too, and that changes what you can build with it today.
JEV adds image processing and a new Box primitive
JEV is the typed decision model from TypeSafe AI: you give it content, you give it a question with a fixed answer shape (a choice from a list, a score, a yes/no), and it hands back a structured answer instead of a paragraph you have to parse. Up until now that content was text only. TypeSafe has added image support plus a new question type called Box, which finds where something is in a picture and returns the box edges as fractions of the image's width and height.
A Towards Data Science walkthrough ran it through four use cases: classifying fabric patterns against reference photos, scoring how sharp a product photo is, locating and cropping a shipping label, and spot-the-difference between two similar images. The fabric classification example is the clean one to study: feed JEV three reference images (striped, checked, floral) plus one to classify, and it returned the striped match with a confidence of 1.0 and all other option probabilities at 0.0. Fed a polka-dot image instead, it correctly bucketed that as "other" with the same 1.0 confidence. For the sharpness check, it scored a sharp teapot photo and a blurry one on a 0 to 3 scale for focus alone, rather than mashing lighting, composition and sharpness into one vague number.
There are real limits worth knowing before you build around this: images have to be JPEG or PNG, between 280 and 1568 pixels on each side, under 16 MiB per image, and the whole request body (JSON plus base64 plus framing) has to stay under 32 MiB. You'll also need the typesafe-sdk[pil] package, the jev-1.14.0 model, and a header flag to turn images on.
Why this matters if you're building with AI assistants: a lot of "AI vision" tasks in real apps aren't creative, they're classification with a yes/no or a pick-one answer at the end, like "does this product photo need to go back to the seller" or "where's the shipping label in this frame." That's exactly what JEV's new question types are built for, and you don't have to parse a model's chatty response to get there.
[YOUR TAKE: have you tried routing anything through JEV's image endpoints yet, or still leaning on vision APIs that return free text?]
A local model that does the PC chores you keep putting off
An XDA piece ran Unsloth's quant of Qwen3.6-35B-A3B locally through llama.cpp and Hermes Agent on an RTX 4070 Ti Super, getting 56 to 58 tokens per second. The model used winget to find more than 40 apps with pending updates and finished the picked updates in 1 minute 42 seconds. It set up nightly backups with robocopy and a real Windows scheduled task, so the backup runs at 11 PM even with the model offline, and copied 7 files in 2.3 seconds in testing. It also ran a network diagnostic (speedtest plus pings to the router, 1.1.1.1 and google.com, plus DNS lookups) in 1 minute 44 seconds, and the author used that exact check during the Microsoft outage that hit Outlook and Teams across 74 countries, confirming within two minutes that the problem wasn't on their end.
Widgets are the easiest first Android project to vibe code
How-To Geek makes the case that a home screen widget, not a full app, is the right first project for vibe coding Android with an AI agent in Android Studio. The author's earlier projects, a Wear OS watch face and a full launcher, took days to weeks of back and forth. The widget took under an hour. He fed the agent a mockup image plus a text description of a weather widget, and the first attempt got roughly 70% of the way there visually, just with no working data behind it. One follow-up prompt asking for live weather got the Open-Meteo API wired in and tap-to-refresh working within minutes. The rest was visual polish.
What this means for you
- If you're classifying or scoring images anywhere in your stack (QC photos, shipping labels, before/after comparisons), swap that custom logic for a JEV Choice, Score or Box call this week. Mind the limits: 280 to 1568 pixels, 16 MiB per image, JPEG or PNG.
- Got repetitive PC maintenance tasks eating your attention? A local model through something like llama.cpp plus an agent framework can own app updates, backups and network checks without touching your cloud API credits.
- New to Android and want a real first win instead of a multi-week slog? Build a widget, not a full app. One interface, no app drawer entry, nothing else to design.
- When you brief an AI agent on UI work, pair an image mockup with a plain-text description of behavior. That combo got the Android widget 70% done on the first try.
What's the first thing you'd point JEV's new image support at? Or if you've handed chores to a local model, what's on its to-do list? Drop it in the thread.
How today's stories score
- JEV adds image processing and a new Box primitive: Signal 84 · Facts 54 · Hype 20
- A local AI model takes over Windows chores: Signal 57 · Facts 65 · Hype 17
- Start Android vibe coding with a custom home-screen widget: Signal 54 · Facts 50 · Hype 21
Scored by an AI decision model from each story's articles. Facts means how checkable the sources look, not that we verified them.
Sources
- JEV Sorted Fabric Patterns With 1.0 Confidence and Flagged a Blurry Teapot Photo – Towards Data Science, 2026-11-13
- I gave a local AI model control of my Windows PC, and it handles the chores I hate – XDA, 2026-11-13
- A custom home screen widget is the easiest way to start vibe coding Android apps – How-To Geek, 2026-11-14