Jev 1.14.0: adds experimental image input and Box type

Yeon Vail @yeonvailOriginally posted 2026.10.15

Open original on X

an ex-OpenAI researcher built a model that refuses to talk. it answers in 70–500 ms with a type and a probability. in this week's update it got eyes Jev 1.14.0, released by TypeSafe AI on october 14 2026. the changelog says "most things should be a tiny bit better", mostly multilingual. the real news sits in the "danger zone": images, a Box type and longer context, all experimental most will scroll past a "tiny bit better" changelog. i read it as robot news the idea: robots do not need one brain. they need a slow one that plans and a fast one that flinches i read the release, the docs and the outside tests under the purple lamp in praga what runs under the hood: 1. THE COMPANY: TypeSafe AI, San Francisco, founded 2024. out of stealth september 15 2026 with a $40M seed led by DCVC, ~$200M valuation per Forbes. CEO Diogo Almeida spent ~4 years at OpenAI on InstructGPT, ChatGPT and GPT-4. Jev is named after Jevons: make thinking cheaper and it ends up everywhere 2. THE OUTPUT: no text at all. 3 question types: Choice (pick 1 of N), Score (a point on your scale), Noul (yes/no as a probability). all questions run in 1 parallel pass, answers cannot leave the schema, each comes with a calibrated probability and a confidence 3. THE SPEED: vendor says 70–500 ms. AY Automate's own test on OpenRouter: median 0.33 s, slowest of 791 calls 1.42 s. a developer running it in Browser Use reports 150–300 ms per step 4. THE PRICE: $0.042 per 1M input tokens, output is free. TypeSafe claims up to 193.6x faster and 444.6x cheaper on its own workflows. an independent test by Every on 1 extraction task, as reported by Forkast: ~25x faster, ~580x cheaper than a frontier chat model 5. THE EYES: 1.14 adds images inside state, instructions or criteria behind the header TypeSafe-Danger-Zone-Images: true, plus a Box type that returns 1 bounding box. TypeSafe itself warns multimodal is worse than text and there is "too much jaggedness" for production 6. THE TEST DRIVE: a developer already wired Jev through a harness to play Time Crisis, an arcade shooter, and said it "did pretty well". no numbers. alone each piece is a cheap classifier. together they are a sub-second reflex that can look at a frame and point at a box. that is the shape of a robot's System 1, even though TypeSafe has not announced any robotics use my sketch, not a TypeSafe feature: CAMERA FRAME → JEV: BOX + NOUL "SAFE TO GRAB?" 0.95 → ACT · or LOW CONFIDENCE → ESCALATE TO THE SLOW MODEL THAT PLANS my take: the last two weeks were about slow giants. Astra pauses for seconds, Argon writes simulations, Unitree waits for GPT-7. Jev is the opposite bet: small, typed, cheap, fast. but be honest about the speed: 70–500 ms is ~2–14 decisions a second, and joint control runs at 50 Hz, one command every 20 ms. this is a reflex for decisions, not for muscles the novelty is not that a model got cheaper. the novelty is a lab that is proud its model cannot talk, and whose next step was to give it eyes bookmark this and watch whether the Box type leaves the danger zone edge is open until someone wires a System One into a robot arm and posts the latency

This post quotes another post: x.com/lctrnotes/status/2110645636148105676