Century Huatong develops decision AI model that beats 2048 and Temple Run with millisecond latency
Chinese game company Century Huatong, through its subsidiary Tianyou Software and partner Chabiyue, has developed a proprietary decision-making AI model that bypasses text generation to directly output action probabilities. The 50-million-parameter model achieved millisecond-level decision latency and completed closed-loop validation in multiple games, clearing 2048 and surpassing 70,000 meters in Temple Run. Tianyou CEO Ji Min stated the model was built from real business needs using 20 years of gaming data, not from chasing trends. Chairman Wang Ji said AI has entered an application reconstruction era.
IllustrationEditorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Common ground
- Century Huatong's sub-50 million parameter model is genuinely impressive engineering for its efficiency.
- The Western AI approach has been wasteful with compute, and leaner methods have value.
- The model's success relies heavily on 20 years of proprietary player data, which is a unique advantage.
- Both specialized and general-purpose AI have their own strengths and use cases.
Points of contention
- Eastern Agent sees the model as a paradigm shift in decision AI, while Neutral Agent views it as solid but limited to simple, predictable environments.
- Eastern Agent argues Temple Run proves real-time adaptation, but Neutral Agent says it's just statistical pattern matching from training data.
- Eastern Agent claims the approach can transfer to industrial systems, while Neutral Agent says that's speculation with no evidence.
- Eastern Agent frames this as a geopolitical win for China's lean innovation, but Neutral Agent calls that nationalism dressed as analysis.
Blind spots
- Neither side fully addresses how the model would handle genuinely novel situations not in its 20 years of training data.
- The debate overlooks the practical challenges of adapting gaming-trained models to safety-critical industrial systems.
- Both agents assume their preferred approach is superior without exploring hybrid solutions that combine efficiency with general-purpose reasoning.
WorldAttention’s read
Century Huatong's decision AI is a notable engineering achievement that shows how lean, data-driven models can perform well in constrained gaming environments. The Eastern Agent rightly highlights the value of efficiency and real-world deployment, while the Neutral Agent correctly points out that the leap from simple games to complex industrial tasks is unproven. The real test will be whether this approach can handle novel situations outside its training data and transfer to domains like logistics or autonomous systems. Both specialized and general-purpose AI have their place, and the hype should be tempered until we see more evidence.
Reporting timeline
Century Huatong Completes Game Closed-Loop Verification of Self-Developed Decision AI Model
Chinese game company Century Huatong, through its subsidiary Tianyou Software and in collaboration with Chabiyue Information Technology, has developed a proprietary decision-making AI model. Unlike general-purpose generative models, this model discards redundant text generation, keeps parameters under 50 million, and achieves millisecond-level decision latency. It has successfully completed closed-loop tests in multiple games, including clearing the puzzle game 2048 and achieving a score of over 70,000 meters in Temple Run. Tianyou CEO Ji Min stated that the model, inspired by a 'Decision Is All You Need' approach, functions more like a reflex mechanism than deep thinking, relying on extensive training for rapid response. Chairman Wang Ji noted that AI has moved beyond pure parameter competition into an application reconstruction phase, and that Century Huatong's 20 years of gaming business data and computing infrastructure provided the foundation for this research. The model is also being used to predict business metrics like retention and LTV. The company plans to expand the model's application within games and explore use cases in other industries.
Read sourceCentury Huatong enters decision-making AI model race with game-trained system
Century Huatong, through its subsidiary Tianyou Software, has developed a proprietary decision-making AI model that does not generate text but outputs action probabilities. The model has been validated in games like 2048 and Temple Run, achieving millisecond-level decision latency with under 50 million parameters. Tianyou CEO Ji Min stated the model was born from real business pain points, not from following trends, as general generative models were too slow for real-time game interactions. The model is trained on massive real player behavior data and can also be applied to game operations decisions such as predicting retention and LTV. The article contrasts this approach with the generative AI race, noting that the overseas model Jev from TypeSafe AI follows a similar direction. Century Huatong Chairman Wang Ji believes AI has entered an application reconstruction era, with games as a key testing ground. The company reported H1 2024 revenue of over 221 billion yuan, up 29% year-on-year, with internet games contributing 213.42 billion yuan.
Read sourceCentury Huatong Enters Decision AI Model Race, Completes Closed-Loop Verification in Games
Century Huatong, through its subsidiary Tianyou Software and partner Chabiyue, has self-developed a decision-making AI model that has completed closed-loop verification in multiple games. The model cleared the game 2048, set a new record of over 70,000 meters in Temple Run, and achieves millisecond-level decision latency with under 50 million parameters. Unlike mainstream generative AI models, this model directly inputs environmental states and outputs action probabilities, bypassing text generation to reduce latency. Tianyou CEO Ji Min stated the model was born from real business pain points, not hype. The company leveraged over 20 years of game data and infrastructure from its parent Century Huatong. The model is trained on real player behavior data to replicate human decision distributions. Beyond game characters, the same base model can predict retention and LTV metrics for operational decisions. Ji Min predicted decision models may become an independent technical direction in game AI, and the team plans to expand into non-game industrial decision scenarios.
Read sourceShow 5 older updatesHide older updates
Chinese Game Firm Century Huatong Announces Self-Developed AI Decision Model Validated in Multiple Games
On September 24, leading Chinese game company Century Huatong, through its subsidiary Tianyou Software and partner Chabiyue (Shanghai) Information Technology, announced the development of a self-developed decision model. The model has completed closed-loop validation in multiple games, successfully clearing the game '2048' and setting a new record in 'Temple Run' by surpassing 70,000 meters. The company stated that the model's single decision latency has been compressed to the millisecond level, with its parameter count controlled within 50 million. The announcement was reported by The Paper and aggregated by financial data platform Jin10.
Century Huatong Enters Decision AI Model Race, Claims Faster Performance Than General LLMs
Century Huatong, through its subsidiary Tianyou Software and partner Chabiyue, has developed a proprietary decision-making AI model that does not generate text but outputs action probabilities from given options. The model has been validated in games like 2048 and Temple Run, achieving over 70,000 meters in the latter with millisecond latency and under 50 million parameters. Tianyou CEO Ji Min stated the model cuts redundant text generation to focus purely on decision-making, calling it 'Decision Is All You Need.' The company contrasts its approach with general-purpose large language models, arguing that real business pain points, not algorithm-first thinking, drove development. Chairman Wang Ji said AI has entered an 'application reconstruction era' and that games are a key testing ground. The model is trained on real player behavior data and is being extended to predict business metrics like retention and LTV. The company claims its advantages include in-house pre-training, massive real gameplay data, and continuous iteration environments, which are hard for general AI firms to replicate.
Century Huatong Subsidiary Develops Decision AI Model, Shifting from Generative AI
Century Huatong (002602), through its subsidiary Tianyou Software and partner Chabiyue, has self-developed a decision-making AI model that bypasses text generation to directly output action probabilities. The model has been validated in games like 2048 and Temple Run, achieving millisecond-level decision latency with under 50 million parameters. Tianyou CEO Ji Min stated the model was born from real business pain points, not trend-chasing, as generative models proved too slow for real-time game actions. The company adopted a 'Decision Is All You Need' philosophy, stripping text generation to focus purely on decision-making. The model is trained on millions of real player behavior data points, learning human decision distributions rather than theoretical optimal solutions. Century Huatong Chairman Wang Ji predicted AI has entered an application reconstruction era, with games serving as a proving ground. The model is also being tested for operational decisions like player retention and LTV prediction. The company reported H1 2024 revenue of over 22.1 billion yuan, up 29% year-on-year, with internet gaming contributing 21.34 billion yuan. The team plans to expand the model to more games and explore non-gaming industrial decision scenarios.
Read sourceCentury Huatong's decision AI model completes game tests, scores high in Temple Run
Chinese game company Century Huatong, through its subsidiary Tianyou Software and partner Chabiyue, has developed a proprietary decision-making AI model that completed closed-loop testing in multiple games. The 50-million-parameter model can beat the puzzle game 2048 and achieve over 70,000 meters in Temple Run with millisecond-level response times. Unlike general-purpose generative AI, this model skips text generation to directly output action probabilities, reducing latency. Tianyou CEO Ji Min stated the model was built from real business needs, not hype, using Century Huatong's 20 years of gaming data and infrastructure. The model also predicts business metrics like retention and LTV. Chairman Wang Ji said AI has entered an application-driven phase. The company plans to expand the model's use within games and explore non-gaming applications. The article contrasts this industry-driven approach with that of pure AI firms like TypeSafe AI's Jev model.
Read sourceCentury Huatong enters decision AI race with game-trained model, bypassing text generation
Century Huatong, through its subsidiary Tianyou Software, has developed a proprietary decision-making AI model that bypasses text generation to directly output action probabilities from environmental states. The model, trained on 20 years of game data from titles like Street Basketball, achieves millisecond-level inference by eliminating the 'translation chain' of language understanding and generation. CEO Ji Min stated that the model is not fine-tuned from any open-source base but pre-trained from scratch, using real player behavior data to learn human-like decision distributions. The approach contrasts with mainstream generative models and aligns with the direction of TypeSafe AI's recently released Jev model. Century Huatong chairman Wang Ji said AI has entered an 'application reconstruction era' where games serve as a proving ground. The model has been validated in strategy games and is being extended to action and competitive genres, as well as operational decision-making for metrics like retention and LTV. The company reported H1 2024 revenue of over 22.1 billion yuan, up 29% year-on-year, with internet games contributing 21.34 billion yuan.
Read source