How China’s AI Labs Are Closing the Gap with the US in Q2 2025

Report Overview

The Artificial Analysis team released the State of AI: China Q2 2025 Highlights Report , which benchmarks Chinese AI development using the Artificial Analysis Intelligence Index. Seven evaluation metrics—MMLU‑Pro, GPQA Diamond, Humanity’s Last Exam, LiveCodeBench, SciCode, AIME, and MATH‑500—are combined to assess model performance and real‑world use cases.

China‑US Frontier Gap Shrinks

Since ChatGPT’s launch in late 2022, the performance gap between Chinese and US frontier language models has narrowed from over a year to less than three months. DeepSeek R1 (released May 2025) leads Chinese labs, while OpenAI’s o3 remains the top US model. Chinese progress is driven mainly by DeepSeek and Alibaba; the US relies on OpenAI.

Open‑Source Weight Models

In November 2024, Alibaba’s QwQ 32B Preview surpassed Meta’s Llama 3.1 405B. DeepSeek’s R1 (Jan 2025) was the first open‑weight model to compete with OpenAI’s o1, and R1‑0528 (May 2025) is currently the most intelligent open‑weight model. Chinese labs favor flagship open‑weight releases, contrasting with the US’s more closed‑source strategy.

Steady Advancement of Leading Chinese Labs

DeepSeek and Alibaba dominate Chinese AI. By May 2025, DeepSeek R1‑0528 slightly outperforms Alibaba’s Qwen3 235B A22B . Both companies publish new models roughly every three months, using open‑weight policies to encourage broad adoption.

DeepSeek’s model intelligence has risen from DeepSeek LLM 67B (score 20) in Nov 2023 to R1‑0528 (score 68), with intermediate versions V2, V2.5, V3, and R1. Reinforcement learning (RL) updates have been crucial, positioning DeepSeek as the world’s second‑most capable AI lab alongside xAI, Meta, and Anthropic.

US Competition Intensifies

OpenAI’s dominance is waning as Google, xAI, and Anthropic close the gap. As of May 2025, OpenAI o3 remains the most intelligent US model, followed by Google’s Gemini 2.5 Pro , xAI’s Grok3 mini reasoning (high) , and Anthropic’s Claude Opus 4 (Extended Thinking) .

Classification of Chinese AI Players

Chinese AI entities are grouped into three categories:

Big Tech Companies : Alibaba (Qwen series, Tongyi Qianwen), ByteDance (Doubao, Seed‑Thinking‑v1.5), Huawei (Pangu 5.0, Celia), Tencent (Hunyuan TurboS, Yuanbao, Yuanqi), Baidu (ERNIE 4.5, ERNIE X1, Wenxin Yiyan).

AI Start‑ups : DeepSeek (V3, R1), Moonshot AI (v1, Kimi K1.5), Zhipu (GLM‑4‑32B, GLM‑Z1‑32B), StepFun (Step‑2, Step‑R1‑V‑Mini), MiniMax (MiniMax‑Text‑01, Talkie AI), 01.AI (Yi‑Lightning, YiChat), Baichuan (Baichuan 4, Baichuan M1).

Other Ambitious Companies : Kunlun Tech, 360 Security, iFlytek, Meituan, Xiaomi.

Leading Language Models and Open‑Source Frontier

Top Chinese models include DeepSeek R1 (May 2025, score 68), Alibaba Qwen3 235B A22B (Reasoning, score 62), and ByteDance Seed‑Thinking‑v1.5 (score 62). US leaders are OpenAI o3 (score 70), Google Gemini 2.5 Pro (68), and xAI Grok 3 Mini Reasoning (67).

Open‑source weight leadership is held by DeepSeek (R1 May 2025, score 68; V3, score 52). Chinese models dominate both inference (e.g., Qwen3 235B A22B Reasoning, score 62) and non‑inference rankings.

Multimodal AI and Media Generation Progress

Chinese firms are active across text, speech, image, video, and 3D generation. Alibaba contributes image (LHM) and video (Wan 2.1); ByteDance offers TTS (Seed‑TTS) and video (Seaweed‑7B). In text‑to‑image, US and China have reached parity, with OpenAI GPT‑4o (ELO 1165) slightly ahead of ByteDance Seedream 3.0 (ELO 1161). In text‑to‑video, Google Veo 3 Preview (ELO 1247) leads, followed by Kuaishou Kling 2.0 (ELO 1133) and China’s MiniMax T2V‑01 (ELO 1053) and Alibaba Wan 2.1 (ELO 1039). Image‑to‑video remains US‑led, with Google Veo 3 (ELO 1222) marginally ahead of Kuaishou Kling 2.0 (ELO 1206) and Runway Gen 4 (ELO 1199).

Conclusion

The report shows that China’s AI ecosystem is booming, especially in open‑source and multimodal domains, and is rapidly narrowing the performance gap with the United States.