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

Forecast: Simultaneous Chinese leadership of the AA and Vals indices

Executive Summary

Probability that the question resolves YES: 24.8%. Simultaneous Chinese leadership is a meaningful minority outcome, but it is not the most likely result.

The requirement is strict: one model developed by an organization headquartered in the People’s Republic of China must hold sole or officially joint first place on both overall public leaderboards during an overlapping period between July 22, 2026 and December 31, 2028, inclusive. A brief overlap counts; different Chinese models leading the two boards does not.

The latest rankings show no Chinese leader on either index, and the historical review found no verified qualifying overlap. Chinese models have nevertheless come close enough to make a future success credible. The main obstacle is achieving broad leadership against an advancing competitive frontier—and having that leadership recognized on both boards at once.

Key drivers

1. The current gaps are substantial on both boards

The latest recorded standings identify Kimi K3.1 Pro as the strongest scored Chinese contender:

Sources: Artificial Analysis leaderboard and Vals leaderboard.

These are different scoring scales, not interchangeable measures of intelligence. AA also lists multiple evaluation settings for individual models, so row rankings overstate the number of distinct competitors ahead of Kimi.

2. Broader testing weakened the apparent catch-up case

AA’s October 27, 2026 methodology refresh was the most important downward update. Kimi K3.1 Pro’s displayed deficit against the same incumbent, Claude Opus 5.5, widened from three points under the previous methodology to twelve under version 5.

That was a change in measurement, not deterioration in the model. It showed that the earlier aggregate concealed important weaknesses. Kimi led ordinary Terminal-Bench 4.0, a computer-task evaluation, while trailing substantially on scientific terminal tasks. Strong coding performance alone is therefore insufficient for overall leadership.

The evidence does not establish that Chinese developers merely optimized familiar public tests. Task selection, difficulty and the use of private tests changed together. See the AA version 5 results and methodology.

3. Genuine near misses keep success plausible

In the July 27 Vals snapshot, Kimi K3 scored 74.70% against the leader’s 75.14%—only 0.44 percentage points short, although explicitly not first. Kimi K3.1 Pro subsequently launched only three displayed AA points behind first under the then-current methodology.

These results demonstrate credible challenges to the overall frontier, rather than leadership only within a Chinese-model or publicly downloadable-model category. Moonshot’s substantial gains from additional training after the base model was built also provide a concrete route to further improvement.

4. Resources favor continued U.S. advantage, but do not fix the outcome

Epoch’s October 27 chip analysis estimates a large U.S. advantage in AI hardware arrivals through June 2026. This supports continued Chinese disadvantage, although hardware arrivals are not the same as operational training capacity or a ceiling on model quality.

There are mechanisms for that disadvantage to weaken. Alibaba said Qwen 4 was in training; Huawei announced deployment of a large Atlas 950 cluster; and Moonshot described substantially more intensive additional training for K3.1 Pro. These are evidence of sustained investment and technical effort—not independent proof that the resulting systems will lead either index.

5. Winning both boards is harder than winning either one

The indices measure related but different capabilities, and their current winners differ. AA version 5 balances knowledge work, coding, science and general capabilities. Vals places most of its weight on finance and coding; U.S. legal and tax content is a smaller part of its aggregate.

A Chinese model could consequently win one board while remaining behind on the other. Evaluation and publication delays could also prevent overlapping leadership. The forecast treats the boards as related, but not interchangeable or statistically independent.

Scenarios

The forecast weighs three competitive regimes. Their weights are judgment-based, rather than frequencies fitted from past dual-leadership events.

The central distinction is between catching up on selected tasks and overtaking the contemporaneous leaders across both evaluation suites. Competitors continue improving throughout the window.

Implications

• Treat dual leadership as a material contingency, not the baseline. Plans premised on permanent U.S. benchmark leadership would overlook a credible alternative; plans assuming inevitable Chinese catch-up would also go beyond the evidence.
• Do not equate a NO outcome with weak Chinese models. A model can be highly capable in coding or financial workflows without simultaneously leading both broad aggregates. Task-specific performance remains more informative for deployment decisions.
• Do not equate a YES outcome with durable dominance. The question can resolve on one brief overlap. That would establish a qualifying benchmark achievement, not sustained superiority or reversal of the resource imbalance.

Signs to watch

The hardware timing comes from Huawei’s September 17 announcement. The capability signals draw on Alibaba’s roadmap and Moonshot’s K3.1 Pro announcement.

Research and methodology

The most consequential sources and original analysis were:

• Official AA rankings and methodology announcements. The current leaderboard, October 23 Kimi assessment and October 27 version 5 announcement established the present gap and the effect of changing the evaluation suite.
• Official Vals rankings and historical captures. The Vals Index supplied scores and weighting rules. Historical reconstruction recovered twelve post-creation captures of the main leaderboard; none showed a Chinese overall leader. These were observations at particular times, not continuous coverage.
• Epoch capability data. Original analysis reconstructed the running capability frontier from the October 30 vintage of Epoch’s Capabilities Index data. None of the 73 China-labelled entries in the 273-entry scored sample reached the contemporaneous sample maximum. This is a separate, retrospectively scored capability measure—not historical proof about AA or Vals listings.
• Hardware estimates and developer disclosures. Epoch’s chip analysis informed the resource constraint. Alibaba, Huawei and Moonshot disclosures supplied evidence of efforts that could weaken it, while remaining unverified indicators of future competitive performance.

The probability comes from a weighted mixture of the three scenarios above. Within each, the calculation distinguishes Chinese AA leadership from the additional requirement of same-model, simultaneous Vals leadership. It allows for repeated opportunities without treating every model update as an independent attempt, and includes small adjustments for missed historical success and publication ending before success. These inputs are judgments, not measured event frequencies.

Limitations

• Historical coverage is incomplete. No uninterrupted record of both boards was recovered. A short overlap between captures cannot be conclusively excluded. Some maintained historical pages also mix old date labels with later results, making actual archive timestamps important.
• The cause and durability of the capability gap remain uncertain. A matched component-level comparison across benchmark revisions would help distinguish persistent broad weaknesses from remediable training gaps and changed task selection.
• Announced hardware is not verified research capacity. Evidence that new systems are reliably supporting successful large training runs would materially improve the assessment.
• Numerical precision exceeds evidential precision. Some current feeds lack exact intraday publication timestamps, and scenario weights remain judgment-based. If a leaderboard ceases publication without a successor before any documented success, the question may be annulled under its rules; cessation is not automatically a NO.