EXPLORINEER
en

DE — heise online

Report: China trains AI by distilling US models

AI1 min read

A report says Chinese AI companies and research labs often train their models by copying answers from big US AI models. This is called distillation, and it helps them work around a shortage of computing power.

What happened

Reuters looked through more than 300 Chinese research papers, patents, company statements, job ads and government documents. It found that Chinese AI firms and research groups seem to use large US models to train their own models. The practice is moving from universities into industry, and the Chinese government supports it so that AI is cheaper for factories to use.

What is distillation?

In distillation, a strong 'teacher' AI model gives answers or step-by-step solutions. A smaller 'student' AI model learns from these answers. The method is not illegal in itself. US AI companies such as Anthropic and OpenAI say Chinese firms break their terms of use. They claim this is done with fake identities, stolen API keys and stolen credit cards. China rejects the accusations and says US companies do similar things.

Key facts

  • Reuters reviewed more than 300 documents from China.
  • Researchers at Shanghai Jiao Tong University picked about 5,000 answers from OpenAI's o1 model.
  • They built 10,750 bilingual training examples and used them to train Alibaba's Qwen2.5-72B.
  • A government plan called 'AI+Manufacturing', published in January by eight agencies, names distillation as a way to support industrial use.

Why it matters

Distillation cannot copy everything a big model can do, and it does not replace strong computer chips. But it can teach small models certain skills. US export bans leave China with few powerful AI chips, so this method helps Chinese firms get more from what they have. It also raises questions about fair use, security and how much China can build its own new ideas.

What to watch

  • Whether US companies or agencies take legal or policy action against Chinese distillation.
  • Whether China gives clear rules on where normal use of foreign models ends.
  • Whether Chinese firms find more computing power outside China to avoid export controls.

Glossary

Reuters
A global news agency that reports news for newspapers, TV and websites.
Anthropic
A US AI company that makes the Claude AI models.
OpenAI
A US AI company that makes ChatGPT and the GPT models.
Alibaba
A large Chinese technology company that builds the Qwen AI models.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

AI전자신문

Lablup Posts 48% Year-on-Year Revenue Growth for Nine Months, First Earnings Report Since Listing

Lablup reported cumulative consolidated revenue of 6.99 billion won and operating profit of 900 million won for the first three quarters, up 48.2% and 36.5% year on year, driven by corporate demand for AI infrastructure. It posted a cumulative net loss of 2.4 billion won due to a 3.49 billion won derivative liability valuation loss tied to the conversion of RCPS shares, but returned to a net profit of 360 million won in the third quarter as that accounting effect expired.

AIZDNet Korea

Maum AI Signs Physical AI Industry-Academia MOU with Hanyang University ERICA and Doosan Robotics

Maum AI signed a three-way MOU with Hanyang University ERICA and Doosan Robotics at the "AI Symposia" held on the 18th at Hanyang University ERICA to build cooperation on physical AI talent development and industry-academia collaboration. The partnership is based on the AX Graduate School project, a government-backed AI talent program run by the Ministry of Science and ICT, with cooperation planned in joint curricula, industry projects, internships, joint R&D and demonstrations, and government R&D planning. Maum AI will contribute its generative and multimodal AI and AI platform capabilities to the effort.

AIheise online

Google Genie 4: Interactive AI Worlds Now Available to Subscribers

Google is rolling out its Genie 4 AI world model to Google AI Ultra subscribers via the Project Genie web app, with access also available in Germany. Users of the cheaper Google AI Pro plan can join a waitlist, and daily usage limits apply. Genie 4 generates interactive 3D environments from text or images, and the model currently supports only about one minute of continuous interaction.

AIheise online

Gemini 4 Neon Challenges Google's Flagship Model

Google DeepMind has released Gemini 4 Neon, the fourth model in the Gemini 4 family within seven weeks, designed for efficiency while matching and slightly outperforming flagship model Carbon on some benchmarks. Neon is slated to become cheaper per token than Carbon from December, when Carbon's introductory prices expire. Google positions it for complex programming, analytical reasoning and enterprise use, filling the gap between Carbon and the everyday model Hydrogen.