TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, released Jev on Sept. 15. The model does not generate text. It makes judgments.
Two weeks later, the rest of the industry was moving in the same direction. OpenAI introduced its Decisions API at DevDay, Cloudflare open-sourced Clef on Oct. 1, and Amazon released Strands Decider. In China, Shanghai AI Laboratory open-sourced Intern-Decision, and on Sept. 30, Shanghai-based startup StartLux, also known as YuanDian XingHui, released the open-source StartLux-Decision and said it beat Jev in public evaluations.
That pace stands out even in AI. A new category went from first launch to a crowded field in roughly two weeks. StartLux is led by Shanda co-founder Chen Danian.
Jev separates judgment from generation
TypeSafe describes Jev as a “System One model,” borrowing the term from Daniel Kahneman’s Thinking, Fast and Slow. System 1 refers to fast, intuitive thinking, while System 2 handles slower reasoning. Over the past two years, most large-model development has pushed toward the System 2 side, with longer chains of thought and deeper reasoning.
Agent workloads look different in practice. Many of the actions inside an agent loop are not deep reasoning tasks but repeated small decisions: which team should receive a ticket, whether a command should be allowed, which button should be clicked next on a page, or whether a task is actually complete.
Before this, those decisions were usually handled in one of two ways. A large model would do them as part of a text-generation step, with the result parsed afterward, or developers would hard-code rules that often failed on long-tail cases. The first option is slow and expensive. The second lacks flexibility.
Jev answers only questions with predefined options, such as multiple choice, scoring, or yes-or-no prompts, and returns structured outputs with probabilities attached. Its price is $0.042 per million input tokens, with output free.
In effect, Jev turns “judgment” into a standalone layer instead of bundling it inside generation. That makes it a cheap, high-frequency piece of infrastructure.
The market response was immediate. Vercel said that within 24 hours of Jev’s launch, 13% of paid users on its AI Gateway had already used it, double the uptake seen with any previous model launch. TypeSafe named the model after economist William Stanley Jevons, a nod to the Jevons paradox: when something gets cheaper, demand can rise rather than fall.
Jev also has a clear limitation. It is a closed-source hosted model, with no public weights and no local deployment option. Users can only access it through TypeSafe’s own API. That gap became the opening for fast followers.
Chinese teams move quickly, with open source and local deployment at the center
Chinese teams have moved quickly into the Jev-style model category.
Shanghai AI Laboratory’s Intern-Decision comes in 0.8B, 2B, and 4B versions. It focuses on multimodal decision-making, meaning it can interpret images and interfaces before making a choice. The team said its inference speed is 2 to 3 times faster than Jev.
StartLux took a broader approach, releasing five sizes from 0.8B to 27B and providing quantized files aimed directly at local deployment.
According to StartLux, its 27B model scored higher than Jev 1.13 in 31 of 38 tests in Decision Index 0.2.1, with an overall score of 63.88 versus 57.91 for Jev. The team also showed a chess demonstration in which the model won 35 out of 36 games.
Those numbers come with caveats. The article says the results were self-tested by the team using public tools and a Sept. 28 snapshot of the leaderboard. The shelf life of first place in this segment is also short. One day after StartLux’s release, Cloudflare said Clef had taken the lead on the same benchmark. As for chess, the article treats it less as a core battleground for decision models and more as a demo built to attract attention.
The more important point is not who is briefly on top. It is that Chinese teams have largely converged on open source and localization.
That is not accidental. Jev is closed-source and cloud-only, which creates a hurdle for developers in China who are sensitive to data handling or wary of cross-border API calls. Open weights and the ability to run models on local machines fill the space Jev left open.
Another detail stands out. Cloudflare’s Clef was post-trained on Qwen. Amazon’s Strands Decider also uses Qwen as its base, and APUS is running on Qwen as well. In this wave of decision models, a Chinese open-source foundation model has become shared infrastructure for global players.
Why StartLux is betting on local AI
StartLux itself was founded this year. Its CEO is Chen Danian, and its CTO is Dr. Guo Quanwei.
Chen is a familiar figure in China’s internet industry. The article says he wrote the internet billing software ENCounter in 1998, co-founded Shanda with Chen Tianqiao the following year, and later built WiFi Master Key.
StartLux describes itself as a company focused on local models, with the goal of keeping both model capability and data on users’ own devices. It drew attention earlier with StartLux-27B, a model post-trained on Qwen3.6-27B. In a specialized MCP test run by the China Academy of Information and Communications Technology, it scored 39.25, ahead of the 284B-parameter DeepSeek-V4-Flash. The article adds an important qualifier: that benchmark focuses on tool use and does not prove broader superiority in general capability.
In StartLux’s roadmap, “local AI” does not mean a small model that can simply be downloaded to a PC. It means a full personal AI system. A 27B model handles general capabilities, a decision model handles high-frequency judgments, the upper layer includes agents and memory, and the lower layer consists of quantization and inference systems.
Seen through that architecture, decision models matter even more for local use than for cloud deployment. Cloud systems can add more compute. Personal computers face hard limits in VRAM, memory, and power consumption. If every small decision has to call a 27B model, a local agent will struggle to run at all. StartLux said its 4B version, compressed to Q4 quantization, is about 2.71GB while still maintaining 98.3% decision consistency with the original weights.
For a company betting on local AI, a decision model is not just a new product category to chase. It is a necessary component if the full local system is going to work.
The article also argues that Jev did not choose the direction for StartLux. On the timeline, StartLux’s decision-model project started later than Jev. What Jev proved was something else: the market is willing to pay for an architecture that splits intelligence into layers and assigns different jobs to different models. That fits the local-system logic StartLux has been describing.
One more number appears in the story: three days. StartLux attributes that speed to its Auto Research system, which uses AI in data construction, training, evaluation, and failure analysis. Whether that method can be reused across different model categories may say more about the company than any single leaderboard result.
Will decision models become a commodity?
It is still too early to say how defensible this layer will be.
Jev was reportedly developed in stealth for two years. OpenAI built its Decisions API in two weeks on top of its own small models. StartLux did it in three days. Cloudflare and Amazon followed within about half a month. When Jev launched, analysts cited in the article expected large companies to release their own decision models quickly, leaving Jev to fight a price war against “frontier versions of itself.”
A category that can be replicated in two weeks is hard to treat as a durable moat.
For StartLux, the harder test lies elsewhere. Guo Quanwei said long-task stability, exception recovery, context management, and the overall user experience are all more difficult than simply getting a model to run.
According to the plan cited in the article, StartLux’s first public-facing local intelligent experience product may launch before the end of the year.
From time-based internet billing tools to WiFi Master Key, many of Chen Danian’s past products were high-frequency, low-level businesses that depended on scale. Decision models fit that pattern too. The open question is where value will settle when judgment becomes cheap enough for everyone: in the model itself, or in the hands of the companies that assemble those models into products.