日本語

MODELS

Models: months behind, not years

If compute were everything, China would be years behind. It is not. We compare the best model of each country on one yardstick, Epoch AI's Capabilities Index (ECI), which combines dozens of tests into a single score.

Key points

  1. China's best model (Kimi K3, Jul 2026) is at a level US models first reached 4.4 months earlier. Over the past year each new Chinese best has been 4–8 months behind.
  2. The gap has neither closed nor opened: it has stayed at a few months since 2024, while the largest US training run used 50x more compute than China's in 2026.
  3. China competes by giving models away: 6 of its top 10 can be downloaded and run by anyone, against 0 of the US top 10.

The best model of each country

Best capability score so far (ECI)Each step up is a new best model. The horizontal distance between the lines is how far behind China is.

Source: Epoch AI (ECI)

Each new Chinese best, and how far behind it was

ModelReleasedECIUS reached itMonths behind
Kimi K3Moonshot · open weightsJul 2026157.7Mar 20264.4
Qwen3.7-MaxAlibabaMay 2026153.7Dec 20255.2
Kimi K2.6Moonshot · open weightsApr 2026151.0Nov 20255.0
GLM-5.1Z.ai (Zhipu AI) · open weightsApr 2026149.9Aug 20258.0
Kimi K2.5Moonshot · open weightsJan 2026148.1Aug 20255.7
DeepSeek-V3.2DeepSeek · open weightsDec 2025146.3Apr 20257.5
Kimi K2 ThinkingMoonshot · open weightsNov 2025146.0Apr 20256.7
DeepSeek-V3.2-ExpDeepSeek · open weightsSep 2025145.0Apr 20255.5
Qwen3-235B-A22B-Thinking (Jul 2025)Alibaba · open weightsJul 2025143.9Mar 20253.8
DeepSeek-R1 (May 2025)DeepSeek · open weightsMay 2025141.3Dec 20245.3
Qwen3-235B-A22BAlibaba · open weightsApr 2025139.4Dec 20244.3
DeepSeek-R1DeepSeek · open weightsJan 2025139.0Dec 20241.1
DeepSeek-V3DeepSeek · open weightsDec 2024132.4Sep 20243.4
Qwen2.5-72BAlibaba · open weightsSep 2024129.0Jun 20243.0
Qwen2.5-32BAlibaba · open weightsSep 2024128.5May 20244.2
Qwen2-72BAlibaba · open weightsJun 2024125.3Mar 202314.8
DeepSeek-V2 (MoE-236B, May 2024)DeepSeek · open weightsMay 2024124.8Mar 202313.8
Yi-34B01.AI · open weightsNov 2023117.3Mar 20237.7
Qwen-14BAlibaba · open weightsSep 2023112.8Mar 20236.4
Baichuan2-13BBaichuan · open weightsSep 2023102.8Feb 20236.4
Baichuan1-7BBaichuan · open weightsJun 202389.8Feb 20233.2

Source: Epoch AI (ECI)

Compute used for the largest training runs

YearUS largestChina largestUS ÷ China
2026GPT-6 Astra1.0×10^27 FLOPKimi K32.0×10^25 FLOP50x
2025Grok 45.0×10^26 FLOPQwen3-Max1.5×10^25 FLOP33x
2024Llama 3.1-405B3.8×10^25 FLOPDoubao-pro2.5×10^25 FLOP2x
2023Gemini 1.0 Ultra5.0×10^25 FLOPQwen-72B1.3×10^24 FLOP38x
2022Minerva (540B)2.7×10^24 FLOPGLM-130B3.5×10^23 FLOP8x
2021FLAN 137B2.0×10^24 FLOPERNIE 3.0 Titan1.0×10^24 FLOP2x
2020GPT-3 175B (davinci)3.1×10^23 FLOPCPM-Large2.6×10^20 FLOP1,205x
2019OpenAI Five6.7×10^22 FLOPMMLSTM (WT-2)1.9×10^17 FLOP345,717x

FLOP = the number of calculations used in training. Estimates by Epoch AI; many recent models do not disclose it, so the largest known run is used.

Source: Epoch AI (models)

How many notable models come from China

Notable AI models by country of the developerChina's share rose from 3% in 2018 to 35% in 2026 (so far).
Show the numbers
YearUSChinaOther
20182719
20195639
20203758
202154169
202268815
2023862112
2024572912
2025643310
202647296

Source: Epoch AI (models)

This page organises information; it does not recommend buying or selling anything. Data belongs to each publisher and may contain errors or delays. Quotes are taken from reports and transcripts; see the sources for the originals. Make your own investment decisions.

Updated 29 Sep 2026 12:24 JST · US power to Jun 2026, China power to May 2026 · models to 9 Sep 2026