technology youtube analysis

China's AI Ascendancy: A Deep Dive into the Players Shaping the Future

With the top open-source AI models now mostly being dominated by Chinese AI labs and even closing in on the performance of private AI models, it’s time for a rundown on Chinese AI developments. This article compiles information on the most popular commercial research labs down to the underground powerhouses making incredible AI progress.

Level 1: The Mainstream - Deepseek

Deepseek is the nonchalant top dog of open-source AI, challenging multi-billion dollar companies with no external funding. They have less compute than major AI labs, yet release most of their work for free, from infrastructure to training scripts. Their most recent paper, “Native Sparse Attention,” won the best paper award at the Association for Computational Linguistics conference. The release of Deepseek R1 even caused a 10% drop in Nvidia’s stock, and the latest Deepseek V3 3.1 now natively supports Chinese hardware.

Level 2: The Chinese Big Three

These are the tech companies with AI budgets that can rival OpenAI or Anthropic.

  • Alibaba Cloud (Qwen): The Chinese version of Amazon, their AI model family, Qwen, is a cornerstone of academic and open-source LLM research. Qwen offers a range of model sizes, making it accessible for various research scales. Their latest model, Qwen 3235B2507 Reasoning, is the number one open-source model.
  • ByteDance (Seed): While late to the AI game, they have shown a strong presence since mid-2024. Their reasoning model, Seed Thinking v1.5, briefly took the crown for state-of-the-art Chinese models. Their image generation model, Seed Dream 3.0, and video generation model, Seedance 1.0, are both number one in their respective arenas.
  • Tencent (Hunren): Tencent’s first LLM, Hunren, was announced in September 2023. They are more secretive with their research but have released open-weight video and text-to-3D models. Hunren is also integrated into WeChat.

Level 3: The AI Startups

  • Moonshot AI (Kimi): Their open-source model, Kimi K2, briefly held the non-reasoning open-weights crown. Founded by a former Meta AI and Google Brain researcher, Moonshot AI was the first to support all 200,000 Chinese characters in their mobile app.
  • Zhipu AI (ZAI): Their GLM4.5 model series entered the top three of the open-source LLM leaderboard. They were also behind the first transformer-based text-to-video generator, CogVideo.
  • Minimax: Known for their AI roleplay chatbot, Glow, and their text-to-speech capabilities, which rival OpenAI and 11 Labs. Their latest model, Minimax M2, has topped the open-weights leaderboard.
  • Kuaishou (Cling): Known for their private video generation model series, Cling, which remains highly rated.

Honorable Mention: Butterfly Effect (Manis AI)

This Singaporean-based company created Manis AI, an autonomous agent application that blew up in March 2025. While the initial hype has died down, it highlighted the growing interest in agentic AI systems.

Level 4: The Pre-Deepseek Era

  • Baidu (Ernie): The Chinese version of Google, their Ernie model series dates back to 2021. They have been mostly closed-source until the release of Ernie 4.5 in July 2025.
  • 01.AI (Yi): Founded by the former head of Google China, they released a near state-of-the-art LLM in November 2023 but have since gone quiet.
  • BRIA: Similar to 01.AI, they released private models in early 2024 but have not been active recently.

Level 5: The Underground

These companies are making significant contributions to AI without gaining mainstream attention.

  • StepFun: The first Chinese AI startup to develop a 1 trillion parameter LLM. Their latest release, Step3, is an open-source multimodal reasoning model.
  • OpenBMB: They contribute to open source by fine-tuning other models with their published datasets. Their MiniCPM model series rivals GPT-4o in language and vision capabilities.
  • Huawei (Pangu): While known for their Ascend chips, their Pangu LLM project has faced setbacks, with a whistleblower exposing that they had been secretly wrapping rival open-source models.
  • SenseTime (SenseNova): Their latest model, SenseNova V6.5, is a multimodal reasoning model that performs slightly better than Gemini 2.5 Pro and Claude for Sonnet. However, their models are not publicly accessible.
  • Shanghai AI Laboratory: A R&D focused program that is a publication machine for top-tier AI research. Their InternLM series is a collection of small open-source models.
  • Ant Group (Ring): An Alibaba-affiliated fintech, they are deep into AI research and have released a one trillion parameter thinking model called Ring 1T.

Honorable Mention: Cherry Studio

An open-source app that connects and integrates LLM providers from within and outside of China, making it easy to try out and switch between models.

Level 6: The Unexpected Players

  • Xiaohongshu (DOTS LM1): The social media platform suddenly released a promising open-source model in June 2025.
  • Xiaomi (Mimo): The phone and car company has stepped into the LLM competition with their open-sourced Mimo model family.
  • Meituan (LongCat-Flash): The Chinese food delivery company dropped a state-of-the-art open-source model, which is like Uber suddenly beating Meta at their own game.

Conclusion

The landscape of open-source AI has significantly shifted towards China. While US companies are slowing down their open-source releases, Chinese companies and institutions are publishing more models and research, benefiting the entire world. This new dynamic may soon shift the power balance in AI development.

This article was written by Gemini, based on content from: https://www.youtube.com/watch?v=XFhUI1fphKU