On September 15, TypeSafe AI, founded by Diogo Almeida, former researcher at OpenAI, launched Jev—a model that writes nothing but makes judgments.

Two weeks later, the entire industry was doing the same thing. OpenAI unveiled the Decisions API during its Developer Day; Cloudflare open-sourced Clef on October 1; Amazon also released Strands Decider.

Domestically, Shanghai AI Lab open-sourced the multimodal decision-making model Intern-Decision.

On September 30, StartLux (formerly Dian Xinghui), a Shanghai-based company founded less than five months prior, announced the open-source release of StartLux-Decision, claiming it outperformed Jev in public benchmarks.

A new category evolved from "first-mover" to "mass competition" within just two weeks—an unprecedented pace even in the large model industry.

The driving force behind StartLux is Chen Dainian, co-founder of Shanda Network.

The Intuition Ignited by Jev

To understand this wave, we must first clarify what decision models actually do.

TypeSafe refers to Jev as a "System One model," drawing its name from Daniel Kahneman’s book *Thinking, Fast and Slow*. System One represents fast, intuitive thinking; System Two handles slow, deliberate cognition. For the past two years, most large models have been pushing toward System Two—longer reasoning chains, deeper thought.

But once Agents truly run in practice, the reality differs. Within an Agent task, frequent occurrences are not deep reasoning, but countless micro-decisions: which team should handle this ticket? Can this command be approved? Which button should be clicked next on the UI? Has the task actually completed?

Previously, such decisions were either delegated to large models “on the side,” requiring text generation followed by parsing; or hardcoded into rules—unable to cover long-tail cases. The former is slow and costly; the latter lacks intelligence.

Jev’s approach is simple: answer only questions with predefined options—single choice, scoring, yes/no—with direct structured output including probability scores. Pricing is $0.042 per million input tokens, with free outputs.

In essence, Jev decouples "judgment" from "generation," turning it into a low-cost, high-frequency infrastructure.

The market response was immediate. Vercel reported that within 24 hours of Jev’s launch, 13% of its paid users on the AI Gateway began using it—double the uptake seen with any previous model release. TypeSafe named the model after economist William Stanley Jevons, alluding to his famous paradox: the cheaper something becomes, the greater the demand.

However, Jev has one glaring gap: it is a closed-source, hosted model with no public weights and no local deployment option—it can only be accessed via TypeSafe’s own API.

This gap precisely became the entry point for later entrants.

Chinese Teams Rapidly Follow Suit

Domestic teams have swiftly followed suit with Jev-like models.

Shanghai AI Lab’s Intern-Decision open-sourced three variants—0.8B, 2B, and 4B—focused on multimodal capabilities, enabling image and interface understanding before making decisions. The team claims inference speed is 2–3 times faster than Jev.

StartLux, a startup, released five versions ranging from 0.8B to 27B, complete with quantized files, directly targeting local deployment.

According to data published by StartLux, the 27B version scored higher than Jev 1.13 in 31 out of 38 tests on Decision Index 0.2.1, with a total score of 63.88 versus Jev’s 57.91. The team also demonstrated a chess match, winning 35 out of 36 games.

These numbers must be interpreted correctly. They were self-tested using publicly available tools, based on a snapshot of the leaderboard dated September 28. More importantly, “first place” in this race has extremely short shelf life. The day after StartLux’s release, Cloudflare claimed Clef led on the same benchmark. As for chess, it’s never been the core domain of decision models—merely a flashy demonstration.

What matters more than who leads temporarily is that Chinese teams have almost unanimously chosen open-source and local deployment.

This is no coincidence. Jev’s closed-source, cloud-only nature poses a barrier for domestic developers concerned about data sensitivity and cross-border API calls. Open-weight models that can run locally perfectly fill the void left by Jev.

Another detail: Cloudflare’s Clef is based on Qwen fine-tuned post-training; Amazon’s Strands Decider also uses Qwen as its foundation; APUS’s solution similarly runs on Qwen. In this wave of decision model enthusiasm, China’s open-source base has become the shared “skeleton” for global players.

Why Betting on Local?

Returning to StartLux itself.

The company was founded this year. CEO Chen Dainian is a well-known figure in China’s internet industry—he wrote the early internet billing software ENCounter in 1998, co-founded Shanda with Chen Tianqiao the following year, and later developed WiFi Key. CTO Guo Quanwei holds a PhD.

StartLux positions itself as a “local model-first” company, ensuring model capability and data remain on users’ own devices. Its earlier breakthrough came with StartLux-27B, fine-tuned on Qwen3.6-27B, which scored 39.25 in China Academy of Information and Communications Technology’s MCP special test—surpassing DeepSeek-V4-Flash, a 284B-parameter model. Note: This is a specialized test focused on tool calling, not indicative of overall general capability superiority.

In StartLux’s vision, “local AI” isn’t just a small model downloadable to a PC, but an entire personal AI system. The 27B model handles general capabilities; the decision model manages high-frequency judgment; the upper layer consists of Agents and memory; the lower layer includes quantization and inference systems.

When viewed in context, decision models hold even greater significance in local environments than in the cloud. While cloud compute power can be scaled up, personal devices face hard limits on GPU VRAM, RAM, and power consumption. If every micro-judgment requires invoking a 27B model, local Agents would be infeasible. According to StartLux’s data, the 4B variant, quantized down to Q4, occupies ~2.71GB and maintains 98.3% consistency with the original weights in decision accuracy.

For a company betting on local deployment, the decision model isn’t a trendy new category—it’s a critical component determining whether the entire local system can function effectively.

This is the true significance of Jev for StartLux. To be precise, Jev didn’t define StartLux’s direction. Chronologically, StartLux’s decision model project clearly came after Jev. What Jev proved instead is that the market has already voted with real money on the idea: “decomposing intelligence into distinct layers, each performing specialized functions”—a principle that aligns perfectly with StartLux’s core local architecture philosophy.

The “3-day” number also deserves mention. StartLux attributes this speed to its Auto Research system, where AI participates in data construction, training, evaluation, and failure analysis. Whether this methodology can be reused across different model categories will reveal far more about the company’s true depth than any single leaderboard position.

Will the Judgment Layer Become Commodity-Level?

Still, it’s too early to judge how deep the moat of decision models truly is.

Jev reportedly spent two years developing in stealth mode. OpenAI built Decisions API in just two weeks based on its internal small model. StartLux achieved its result in three days. Cloudflare and Amazon also followed within half a month. Early analysts predicted that major tech firms would soon launch their own decision models—meaning Jev would eventually compete against “its own cutting-edge version” in a price war.

A category replicable in two weeks cannot form a sustainable barrier.

For StartLux, the real test lies not in the decision model layer itself. Guo Quanwei himself admits that long-task stability, anomaly recovery, context management, and overall user experience are far more challenging than simply running a model.

Planned for release later this year, StartLux’s first public-facing local intelligent experience version is expected to launch.

From time-based internet billing tools to WiFi Key, many of Chen Dainian’s past products were high-frequency, foundational, and scale-driven. Decision models are likewise a high-frequency, foundational operation.

The question remains: when “judgment” becomes cheap enough for everyone to perform, will value ultimately reside in the models themselves—or in those who assemble these models into usable products?

Author｜Hualin Wuwang, Editor｜Jingyu, StartLux Model Score Exceeds Jev｜Image Source: GitHub, StartLux Chess Match｜Image Source: GitHub