Will a Chinese model hold the AA Intelligence Index and the Vals Index simultaneously at any point between now and the end of 2028?

TL;DR

My estimate is 38.6% that a model from a PRC-headquartered developer holds both #1 slots at the same time before 2028-12-31. Kimi K3 is already a near miss: Vals lists it #2 at 74.70%, while Artificial Analysis lists it at 57.1 behind Claude Fable 5 at 59.9 and GPT-5.6 Sol at 58.9 (Vals home; BenchLM AA mirror). The main bottleneck is taking #1 on Artificial Analysis; once that happens, the Vals condition is much easier to satisfy.

Context

This is a transient-overlap question. A Chinese model does not need to be the durable world leader at the end of 2028; it only needs one public snapshot where a PRC-headquartered developer is listed as sole or joint #1 on both target leaderboards. Moonshot AI qualifies under the question definition because its official company page gives a Beijing, China address (Moonshot AI about).

The two target indices are related but not identical. Artificial Analysis Intelligence Index v4.1 is an English, text-only weighted composite of 9 evaluations, with Agents at 34%, Coding at 24%, Scientific Reasoning at 24%, and General at 18% (Artificial Analysis methodology). Vals Index is a finance-and-coding composite using finance weight 2.0 and coding weight 1.4, with private test sets used for published benchmark results (Vals Index; Vals methodology).

Evidence

The historical backbone is mixed. Artificial Analysis says that for most of the period since late 2022, the Intelligence Index frontier was held by one or two labs, but by 2026-07-17 six labs had a model above 50, including two PRC labs: Moonshot AI with Kimi K3 and Z AI with GLM-5.2 (Artificial Analysis frontier note). That makes the old base rate for a PRC all-model #1 low, but the current reference class is no longer the old one. Stanford HAI's 2026 AI Index also says the U.S.-China model-performance gap had effectively closed, with a 2.7% lead for the top U.S. model as of March 2026 (Stanford 2026 AI Index).

The current target-leaderboard state is a near miss, not a YES:

Kimi K3 is the live crux. Artificial Analysis says Kimi K3 scores 57 on the Intelligence Index, remains behind Fable 5 and GPT-5.6 Sol, reaches 1668 Elo on GDPval-AA v2, and takes #1 on AutomationBench-AA (Artificial Analysis Kimi K3 note). Artificial Analysis also says Kimi K3 is second only to Fable 5 on AA-Briefcase, scoring 1543 Elo versus Fable 5 at 1574, on a private long-horizon knowledge-work benchmark (AA-Briefcase note). Vals reports the same model at 74.70% on Vals Index, with 95.10% on SWE-bench Verified, 91.27% on Vibe Code Bench, 80.90% on Terminal-Bench 2.1, 72.61% on CorpFin v2, and 55.88% on Finance Agent v2 (Vals home). This is why I treat Vals as the easier half of the conjunction.

The counterweight is that broad and held-out signals still favor U.S. labs at the absolute top. LM Arena's live text board has Kimi K3 at #10 with Elo 1486, behind Claude Fable 5 at #1 with Elo 1507, while Qwen3.7 Max Preview is #19 and GLM-5.2 is #30 (LM Arena text leaderboard). NIST CAISI reported on 2026-05-01 that DeepSeek V4 Pro was the most capable PRC model it had evaluated, but that it lagged leading U.S. models by about eight months on CAISI's aggregate capability measure (NIST CAISI DeepSeek V4 Pro evaluation). That does not directly rank Kimi K3, but it argues against treating public benchmark closeness as full frontier parity.

Release cadence pushes the other way. Kimi K3 moved Moonshot from Kimi K2.6's much lower level to near-frontier status in one release, and the same Artificial Analysis note says Kimi K3 used 132M output tokens across the nine Intelligence Index evaluations versus 166M for Kimi K2.6 while scoring higher (Artificial Analysis Kimi K3 note). Alibaba has also previewed Qwen3.8-Max-Preview and claimed it is second only to Claude Fable 5, but the reports also say the claim had not yet been independently verified (SCMP on Qwen3.8; SiliconANGLE on Qwen3.8). I give that signal modest weight, not leaderboard weight.

Compute and access keep the estimate below 50%. AP reported on 2026-07-20 that Moonshot paused new Kimi subscriptions after demand overwhelmed capacity, which matters because the target leaderboards require reliable model access for evaluation and continued listing (AP on Kimi capacity). U.S. semiconductor policy is no longer a total wall for all advanced chips, but BIS said on 2026-01-13 that H200, MI325X, and similar exports to China would be reviewed case by case under security requirements (BIS semiconductor export policy). I read this as a drag on sustained Chinese dominance, not a block on a brief leaderboard crossing.

My quantitative model is:

$$P(YES)=P(A) \times P(V \mid A) \times P(R)$$

Here, A means at least one PRC model becomes sole or joint #1 on Artificial Analysis before 2028-12-31, V means a PRC model is also #1 or joint #1 on Vals during the same public-leaderboard state, and R means the leaderboards remain resolvable enough for the question to count. I set P(A) at 58% because the current AA gap is only 2.8 points and two PRC labs are already above 50, but U.S. labs still hold the top two AA positions and have stronger compute depth. I set P(V | A) at 69% because Kimi K3 is already only 0.44 percentage points behind on Vals and the two indices overlap in agentic and coding tasks, but Vals is private and can reward a different profile. I set P(R) at 96.5% because both sources are active and public, with small method-change, publication, and archive risk. The product is 38.6%.

A period-hazard cross-check gives almost the same answer. I use an 11.5% event hazard for the rest of 2026, a 17.5% hazard for 2027, and a 15.5% hazard for 2028. Those hazards reflect one near-term Kimi/Qwen/DeepSeek window, a full 2027 release cycle, and a 2028 cycle where U.S. labs also compound. The combined probability is 38.3%, so I keep the factor-model estimate of 38.6%.

What's non-obvious

The obvious reading is that this asks whether China becomes the global AI leader. That is too broad. The actual event is narrower and more volatile: one Chinese release needs one overlapping public-leaderboard window, and Vals is already close enough that Artificial Analysis is the real bottleneck.

The second non-obvious point is that private benchmarks cut both ways. Vals and parts of Artificial Analysis are designed to limit leakage, so Kimi K3's near-top Vals score is more meaningful than a fully public benchmark result (Vals methodology). But the same private, U.S.-finance, English, agentic setup can also expose gaps that public model-card claims miss, as CAISI's DeepSeek evaluation did for an earlier PRC frontier model (NIST CAISI DeepSeek V4 Pro evaluation).

Limitations

I do not have a complete official daily history of either target leaderboard. That matters because the resolution can turn on a brief overlap that later disappears from the live pages; archived evidence may be decisive even if the year-end leader is non-Chinese.

The Vals site is JavaScript-heavy, so I relied on the official Vals home/model text plus a public mirror for the full July 2026 table. The mirror and the official home differ on 37 versus 38 evaluated models because Gemini 3.6 Flash was added on 2026-07-21, but the top three and the Kimi-versus-Fable gap are unchanged (Vals home; BenchLM Vals mirror).

Both leaderboards can change methodology. Artificial Analysis already moved v4.1 toward agentic workloads, and Vals has updated its formula and components in 2026 (Artificial Analysis methodology; Vals Index). That kind of reweighting is not noise; it can be the difference between a near miss and a resolving YES.