Chinese AI Models vs. US Models: How Close Is China?
The best Chinese model trails the best US model by 27 Arena points, down from 156 in Jan 2024. A year ago it was 27.
The signal
- LaunchesDeepSeek V4.1 Flash (10 Sept 2026). Sources disagree on whether its weights are open.
- LaunchesKimi K2.8 Preview (11 Sept 2026), reported by one tracker only.
- Arenaqwen3.8-max (Alibaba, launched 2 Aug 2026) is the best Chinese model on 13 Sept 2026, #14 overall. 2 Chinese models are in the top 20.
qwen3.8-max trails claude-fable-5.1-max by 27 points. Over the past year the gap has stayed between 25 and 46; it was 156 in Jan 2024.
Highest Arena text rating (overall) from a Chinese lab vs a US lab, last snapshot of each month, Jan 2024 – 13 Sept 2026.
Chinese labs = Alibaba (Qwen), DeepSeek, Z.ai, Moonshot, Xiaomi, MiniMax, Tencent, Baidu, ByteDance, StepFun, Meituan, Ant Group, 01.AI, InternLM. The other line is the best model from any non-Chinese lab; on every snapshot since Jan 2024 that model came from a US lab (OpenAI, Anthropic, Google or xAI), so we label it US. Ranking is by point rating; neighbouring models often overlap within confidence intervals. The dataset leaves out some experimental preview models in early 2025, which may overstate the brief Chinese lead. "ChatGPT" is a product; we compare OpenAI's best Arena model.
Show the data (66 rows)
| Series | Date | Value | Note |
|---|---|---|---|
| Best Chinese-lab model | 2024-01-25 | 1034 | qwen-14b-chat (Alibaba) |
| Best Chinese-lab model | 2024-02-15 | 1147 | qwen1.5-72b-chat (Alibaba) |
| Best Chinese-lab model | 2024-03-27 | 1148 | qwen1.5-72b-chat (Alibaba) |
| Best Chinese-lab model | 2024-04-26 | 1153 | qwen1.5-72b-chat (Alibaba) |
| Best Chinese-lab model | 2024-05-27 | 1168 | qwen1.5-110b-chat (Alibaba) |
| Best Chinese-lab model | 2024-06-29 | 1207 | glm-4-0520 (Z.ai) |
| Best Chinese-lab model | 2024-07-31 | 1207 | glm-4-0520 (Z.ai) |
| Best Chinese-lab model | 2024-08-28 | 1207 | glm-4-0520 (Z.ai) |
| Best Chinese-lab model | 2024-09-27 | 1257 | qwen2.5-72b-instruct (Alibaba) |
| Best Chinese-lab model | 2024-10-28 | 1275 | glm-4-plus (Z.ai) |
| Best Chinese-lab model | 2024-11-22 | 1274 | glm-4-plus (Z.ai) |
| Best Chinese-lab model | 2024-12-30 | 1315 | deepseek-v3 (DeepSeek) |
| Best Chinese-lab model | 2025-01-28 | 1358 | deepseek-r1 (DeepSeek) |
| Best Chinese-lab model | 2025-02-27 | 1363 | deepseek-r1 (DeepSeek) |
| Best Chinese-lab model | 2025-03-30 | 1370 | deepseek-v3-0324 (DeepSeek) |
| Best Chinese-lab model | 2025-04-23 | 1373 | deepseek-v3-0324 (DeepSeek) |
| Best Chinese-lab model | 2025-05-22 | 1368 | deepseek-v3-0324 (DeepSeek) |
| Best Chinese-lab model | 2025-06-24 | 1424 | deepseek-r1-0528 (DeepSeek) |
| Best Chinese-lab model | 2025-07-28 | 1426 | deepseek-r1-0528 (DeepSeek) |
| Best Chinese-lab model | 2025-08-28 | 1429 | glm-4.5 (Z.ai) |
| Best Chinese-lab model | 2025-09-30 | 1439 | qwen3-max-preview (Alibaba) |
| Best Chinese-lab model | 2025-10-17 | 1440 | qwen3-max-preview (Alibaba) |
| Best Chinese-lab model | 2025-11-26 | 1442 | glm-4.6 (Z.ai) |
| Best Chinese-lab model | 2025-12-30 | 1442 | ernie-5.0-preview-1203 (Baidu) |
| Best Chinese-lab model | 2026-01-29 | 1453 | ernie-5.0-0110 (Baidu) |
| Best Chinese-lab model | 2026-02-26 | 1465 | dola-seed-2.0-preview (ByteDance) |
| Best Chinese-lab model | 2026-03-31 | 1472 | qwen3.5-max-preview (Alibaba) |
| Best Chinese-lab model | 2026-04-30 | 1472 | ernie-5.1-preview (Baidu) |
| Best Chinese-lab model | 2026-05-27 | 1475 | qwen3.7-max-preview (Alibaba) |
| Best Chinese-lab model | 2026-06-25 | 1475 | qwen3.7-max-preview (Alibaba) |
| Best Chinese-lab model | 2026-07-30 | 1475 | qwen3.7-max-preview (Alibaba) |
| Best Chinese-lab model | 2026-08-27 | 1480 | qwen3.8-max (Alibaba) |
| Best Chinese-lab model | 2026-09-13 | 1481 | qwen3.8-max (Alibaba) |
| Best US-lab model | 2024-01-25 | 1189 | gpt-4-0314 (OpenAI) |
| Best US-lab model | 2024-02-15 | 1254 | gpt-4-1106-preview (OpenAI) |
| Best US-lab model | 2024-03-27 | 1254 | claude-3-opus-20240229 (Anthropic) |
| Best US-lab model | 2024-04-26 | 1257 | gpt-4-turbo-2024-04-09 (OpenAI) |
| Best US-lab model | 2024-05-27 | 1282 | gpt-4o-2024-05-13 (OpenAI) |
| Best US-lab model | 2024-06-29 | 1287 | gpt-4o-2024-05-13 (OpenAI) |
| Best US-lab model | 2024-07-31 | 1286 | gpt-4o-2024-05-13 (OpenAI) |
| Best US-lab model | 2024-08-28 | 1316 | chatgpt-4o-latest (OpenAI) |
| Best US-lab model | 2024-09-27 | 1339 | o1-preview (OpenAI) |
| Best US-lab model | 2024-10-28 | 1335 | o1-preview (OpenAI) |
| Best US-lab model | 2024-11-22 | 1334 | o1-preview (OpenAI) |
| Best US-lab model | 2024-12-30 | 1359 | o1-2024-12-17 (OpenAI) |
| Best US-lab model | 2025-01-28 | 1351 | o1-2024-12-17 (OpenAI) |
| Best US-lab model | 2025-02-27 | 1359 | gemini-2.0-flash-001 (Google) |
| Best US-lab model | 2025-03-30 | 1406 | chatgpt-4o-latest-20250326 (OpenAI) |
| Best US-lab model | 2025-04-23 | 1418 | o3-2025-04-16 (OpenAI) |
| Best US-lab model | 2025-05-22 | 1409 | o3-2025-04-16 (OpenAI) |
| Best US-lab model | 2025-06-24 | 1477 | gemini-2.5-pro (Google) |
| Best US-lab model | 2025-07-28 | 1474 | gemini-2.5-pro (Google) |
| Best US-lab model | 2025-08-28 | 1466 | gemini-2.5-pro (Google) |
| Best US-lab model | 2025-09-30 | 1467 | gemini-2.5-pro (Google) |
| Best US-lab model | 2025-10-17 | 1465 | gemini-2.5-pro (Google) |
| Best US-lab model | 2025-11-26 | 1488 | gemini-3-pro (Google) |
| Best US-lab model | 2025-12-30 | 1484 | gemini-3-pro (Google) |
| Best US-lab model | 2026-01-29 | 1481 | gemini-3-pro (Google) |
| Best US-lab model | 2026-02-26 | 1502 | claude-opus-4-6 (Anthropic) |
| Best US-lab model | 2026-03-31 | 1501 | claude-opus-4-6-thinking (Anthropic) |
| Best US-lab model | 2026-04-30 | 1501 | claude-opus-4-6-thinking (Anthropic) |
| Best US-lab model | 2026-05-27 | 1499 | claude-opus-4-6-thinking (Anthropic) |
| Best US-lab model | 2026-06-25 | 1500 | claude-opus-4-6-thinking (Anthropic) |
| Best US-lab model | 2026-07-30 | 1512 | claude-opus-5-max (Anthropic) |
| Best US-lab model | 2026-08-27 | 1505 | claude-opus-5-max (Anthropic) |
| Best US-lab model | 2026-09-13 | 1508 | claude-fable-5.1-max (Anthropic) |
The Mode read
DeepSeek and Moonshot both released new models in the week to 13 Sept 2026. That brings back an old question: are Chinese AI models catching up with US models?
They closed most of the gap, then stopped. In Jan 2024 the best Chinese model was 156 points behind. In Jan 2025 deepseek-r1 briefly led outright. Over the past year the gap has stayed between 25 and 46 points. Today it is 27, which means the best Chinese model would win about 46% of blind votes. DeepSeek vs ChatGPT: deepseek-v4-pro ranks #43 and OpenAI's best, gpt-5.5-high, ranks #26.
At 46% of votes, quality does not decide the choice between a Chinese and a US model. Buyers decide on other things: where the data is processed, which rules apply, and whom the buyer already trusts. Those are distribution questions. Each side's models sell most easily where its labs already have the cloud contracts and the compliance approvals.
A 27-point gap does not decide the deal. Trust and procurement do.
What we would back: Companies that sell AI into buyers who choose on provenance and compliance: regulated industries, the public sector, cross-border trade.