Century Huatong subsidiary unveils decision AI model, claims faster performance than general LLMs

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has released a decision-making model called Jev. Unlike conventional large language models (LLMs), Jev does not chat or generate text. Instead, it outputs probabilities for given options. According to the company, Jev is two orders of magnitude faster than existing large models on such tasks. Within days, Jev went viral across the global AI community.

Beneath the buzz, a broader judgment is emerging: the first half of the LLM race was about "learning to talk"; the second half will be about "taking action, executing, and making decisions."

A Parallel Path Emerges in China

While Jev sparked discussion overseas, a different technical route has grown organically from industry in China. Century Huatong subsidiary Tianyou Software, in collaboration with Chabiyue (Shanghai) Information Technology Co., Ltd., has independently developed a decision-making model. The model has completed closed-loop validation across multiple games:

• It successfully completed the game 2048.
• In Temple Run, it has set new records, surpassing 70,000 meters.
• Single-decision latency has been reduced to the millisecond level.
• Model parameters are controlled within 50 million.

From Business Pain Points to Model Innovation

While most AI players start with general-purpose generative models—first solving dialogue and image generation, then seeking downstream applications—Century Huatong took the opposite approach. Starting from its gaming business, it targeted behavioral decision models directly.

"Building decision models is not the exclusive domain of pure AI companies," said Ji Min, CEO of Tianyou Software. In his view, the starting point for game companies doing AI is not the model, but the business.

Over the past two years, China's LLM赛道 has flourished, but structural contradictions have emerged:

• Model companies with algorithmic foundations lack real, high-frequency, strong-feedback business scenarios.
• Industry companies with massive scenarios and real user behavior data mostly rely on third-party models, rarely developing their own underlying models.

Both sides are converging on the same scarce resource: real application data for feedback.

A Strategic Bet on Application-Driven AI

Wang Ji, Chairman of Century Huatong, had long anticipated this shift. He believes AI has moved beyond pure technological breakthroughs into an era of application-driven reconstruction. In his view, games are not just games—they are training grounds for cutting-edge technology and are likely to become the largest application category in the AI era.

This conviction drives Tianyou's AI team to invest in underlying model R&D. Leveraging over two decades of gaming industry experience, data accumulation, and long-term investment in computing infrastructure, the team built a solid hardware foundation for pre-training. The result is a model born from real business pain points, not from chasing trends.

Why Generative Models Fell Short

Tianyou Software has operated the Street Basketball IP for over 20 years, making it the top sports competitive game brand in China. Behind the scenes, the R&D team had long experimented with AI across the full game operation chain. In extensive testing, they hit the ceiling of generative models.

Even a small 0.5B parameter model could not keep up with the fast-paced decision-making required in competitive games. The root cause: generative models must first understand language, then organize language, and finally convert it into action. This "translation chain" introduces unavoidable inference latency.

"Decision Is All You Need"

The team made a decisive cut: remove redundant text generation capabilities and focus all parameters on decision-making itself. Their philosophy: "Decision Is All You Need."

• The model no longer takes text as input or outputs text.
• It directly takes environmental states as input and outputs action decision probabilities.
• By eliminating parameter overhead for language, the model size is drastically reduced, achieving millisecond-level inference and lower computational cost.

Ji Min offers an analogy:

"General-purpose LLMs are like the human brain's deep thinking system—they receive information, complete a full reasoning process, and then output results. Our decision model internalizes decision rules, quickly giving the next action based on the current state. It's closer to the human cerebellum and reflex system—turning millions of training iterations into muscle memory, shortening the entire decision chain for rapid, continuous action."

Two Paths, Same Destination

Jev, which went viral overseas, is a product of algorithmic innovation from an AI company. Tianyou Software's model, by contrast, emerged from real pain points in gaming. Though their starting points differ, they converge on the same direction.

"Different routes, same direction," said Ji Min. "Jev's popularity has made one thing clear to the industry: AI doesn't have to chat. Directly making judgments and outputting actions is equally valuable."

Real-World Validation and Data-Driven Training

The model's capabilities have been rigorously tested in gaming environments. The team encodes game states, historical actions, opponent behavior, resource status, and long-term goals into state vectors, training the model to learn selection logic under different scenarios.

A key innovation: the team introduced massive real player behavior data into training. The model learns not only theoretical optimal solutions but also the behavioral distribution of real players. In the same situation, players of different skill levels and styles make different choices—the model learns to replicate real human decision probabilities.

Expanding Capabilities and Applications

The model's capabilities are already being extended:

• In some strategy games, it has generated positive business data feedback.
• Validation in action and competitive games is underway.
• Engineering deployment is accelerating.

Crucially, the model's use is not limited to controlling in-game characters. Using the same pre-training foundation, it can:

• Receive real-time game operation data
• Predict core metrics such as retention and LTV
• Assist operators in formulating strategies

The same underlying model can handle both in-game agent behavior decisions and game business operational decisions.

Three Irreplaceable Conditions for Self-Developed Models

Achieving a model built from scratch requires three conditions, all of which Tianyou Software possesses thanks to its parent company Century Huatong's deep roots in gaming:

1. Ability to pre-train from scratch — the model was not fine-tuned from any open-source base model.
2. Massive real match and player behavior data.
3. Continuously iterable real-world deployment scenarios.

These conditions form a barrier that many general-purpose LLM vendors find difficult to replicate.

"Although we are a business-driven company, Century Huatong is also doing solid foundational research around its business," said Ji Min. He believes that game companies hold real scenarios and massive real data, while LLM iteration depends heavily on feedback from real business data. The boundary between model companies and industry companies is blurring—and that is why the team continues to invest in decision models.

A New Technical Direction for Game AI

While gaming environments have a relatively higher tolerance for errors, this does not mean trial and error is acceptable. Ji Min emphasized that all models undergo large-scale repeated testing before deployment.

"Decision models may become an independent technical direction for game AI," Ji Min predicted. In the future, game AI systems will feature specialized models for dialogue, content generation, and judgment/action—each with its own role, working in coordination.

Compared to the current industry rush toward general-purpose generative LLMs, decision models target a more hardcore future: behavior execution.

Looking Ahead

For Tianyou Software and Century Huatong, this breakthrough in underlying models is not about chasing a fleeting trend, nor will it stop at a single technology release. The next steps include:

• Continuously refining the decision model foundation
• Bringing it into more game products
• Exploring decision-making scenarios beyond gaming

Perhaps, while the rest of the industry is still competing on chat and generation capabilities, Century Huatong's decision model will have already climbed to the next level.

For more exclusive features and expert columns, visit the Finance Channel. Editor: Lishu