Jev Ignites the AI Decision Model Race: Chinese Teams Follow With Open Source and Local Deployment

On September 15, TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, released Jev. This model does not write a single word; it only makes judgments: it answers questions with preset options and directly returns structured results with probabilities for single-choice, scoring, or yes/no questions.^[1]

Jev is priced at $0.042 per million input tokens, with output free. Market reaction came quickly. Vercel disclosed that within 24 hours of Jev's launch, 13% of paid users on its AI Gateway were already using it, twice as many as for any previous model release. TypeSafe named the model after economist Jevons, alluding to the famous paradox: the cheaper something is, the greater demand becomes.^[1]

Behind this is a change in how agents operate. In an agent task, what frequently happens is often not deep reasoning but a large number of trivial judgments: which team should this ticket be assigned to, can this command be allowed, which button should be clicked next on the page, and whether the task is actually complete. In the past, these judgments were either left to large models to do in passing, generating a piece of text and then parsing it, or hardcoded into rules that could not cover the long tail. Jev separates judgment from generation and turns it into cheap, high-frequency infrastructure.^[1]

But Jev has an obvious gap. It is a closed-source hosted model, with no public weights and no local deployment, and can only be called through TypeSafe's own API. This gap became precisely the entry point for later entrants.^[1]

Chinese teams quickly followed. Shanghai AI Laboratory's Intern-Decision open-sourced three tiers, 0.8B, 2B, and 4B, focusing on multimodality and able to understand images and interfaces before making decisions. The team reports inference speeds 2 to 3 times faster than Jev. Startups such as StartLux released five size variants at once, from 0.8B to 27B, complete with quantization files, directly targeting local deployment.^[1]

On September 30, StartLux (原点星辉), a Shanghai company less than five months old, released the open-source StartLux-Decision and claimed it surpassed Jev in public evaluations. The person at the helm behind StartLux is Chen Danian, co-founder of Shanda.^[1]

In the same time window, there was a dense flurry of industry moves. OpenAI launched the Decisions API at Developer Day, Cloudflare open-sourced Clef on October 1, and Amazon also released Strands Decider. A new category went from first release to group battle within two weeks, a speed rarely seen even in the large model industry.^[1]