Decitron bets AI's next act is simulating the world, not answering
Two stories out of China this morning, and they point at the same pressure: the industry has moved past "the model answered correctly" and into "does this hold up in the real world."
Wenge AI has published the full technical report for Decitron, a "general decision model" designed to simulate how a situation evolves rather than answer a question about it. The system, launched in June and detailed in a paper deposited with ChinaXiv on August 3, wraps a world model (MetaWorld), multi-agent simulation, game-theoretic solving and probabilistic forecasting into one loop: current world state, who can act, what happens, repeat. The claim it is making is narrow but real — that forecasting, gaming out rivals and pricing uncertainty should share one continuously updated state, not be three separate prompts.
The numbers are the vendor's own, which is the right caveat here. On a multi-agent game benchmark called TMGBench, Decitron reports 99.4% equilibrium accuracy and 0.8-second average solve time; on PolyBench, 81.2% final-forecast accuracy against named baselines including Gemini 3.8 Flash and Grok 4.1 Fast. The sharpest result is on short-horizon event forecasting, where it scores 72.8% on three-to-30-day questions against 35.4% for GPT-5.5 and 44.6% for Claude 4.7 — and where the gap closes as the horizon stretches.
That last pattern is the interesting part, and also the honest limit. Structured simulation appears to help most exactly where pattern-matching on history helps least: fast-moving situations with no clean precedent. It helps least over long horizons, where the accumulated error in a self-updating world model has more time to compound — if the model's read of today drifts, every downstream branch inherits the drift. What Wenge has actually demonstrated is a plausible architecture, benchmarked by its own authors; nobody has independently reproduced it yet.
Zhipu used its first earnings call as a listed company to introduce a metric the industry has been missing: the "compute multiplier," or how much API revenue each yuan of training and inference compute produces. It pairs with "effective compute" — silicon actually deployed, scheduled and converted into billable tokens, rather than silicon sitting in a rack. Together they are an attempt to answer the question every lab now faces: not how good is the model, but how much revenue does a GPU generate.
The numbers behind the framing are strong. Zhipu reported first-half API gross margin moving from negative 0.4% to 24.6%, average API prices up about 101% since January, and unit inference cost per token down 80% — growth that is not being bought with discounts. Its GLM-5.3-Flash, the first model served entirely on a large domestic chip cluster, pulled 62 trillion tokens in six days.
What to watch: whether anyone outside Wenge reproduces Decitron's short-horizon advantage, and whether other labs start reporting a compute multiplier of their own.
Would you trust a model's simulated future over its answer about today? Tell us in the comments.
