Open AI Models Challenged by Chinese Competition

Western companies are being questioned, some practices are shifting towards Chinese models, the technical gap with closed systems is narrowing, and controversies surrounding distillation are emerging: 2026 is reshaping the landscape of open models. Technical and political sources describe a scenario where security and national strategy weigh as heavily as benchmarks.

Companies Under Scrutiny and Shifts Towards Chinese Models

The use of Chinese models by Western companies has drawn significant regulatory attention. Legislators questioned DoorDash on July 31, 2026, as well as Airbnb and Anysphere/Cursor on April 29, 2026, and Apple on May 17, 2025, regarding their reliance on these systems. Concurrently, technical and economic decisions have led some companies to change their tools: Perplexity quickly adopted DeepSeek R1, according to information from January 28, 2025, and Thomson Reuters has built on Qwen to move away from Claude, according to information from August 24, 2026.

Performance: Gap Narrowed to 4–6 Months and Open Leadership from China

The gap between open and closed models has narrowed to around 4 to 6 months. Since around 2024, the leading open models have come from Chinese laboratories. Independent assessments published in August 2026 conclude that open models are approaching the frontier, while data sources from Epoch AI and Artificial Analysis track the evolution of this gap over time. Another analysis in May 2026 combines public and private evaluations. On November 21, 2025, a product manager from Z.ai claimed the rapid pace of open-source releases, mentioning that they could go live within hours.

Distillation: Trace Extraction Techniques and Controversies in 2026

In 2026, distillation has emerged as the central debate surrounding open models. A chapter from a post-RLHF training manual published in 2026 revisits the fundamentals of synthetic data and distillation, while studies detail how this approach aids Chinese laboratories without diminishing their capacity for innovation. API implementations enabling systematic extraction of reasoning traces were presented in 2026, and Anthropic confirmed that this technique is utilized by Chinese laboratories. Some positions deem the political panic attributing the proximity of Chinese models to the frontier solely to distillation as unfounded, describing current uses and the performance gains observed. In 2024, writings noted the dominance of SFT over outputs from stronger models and the transition initiated towards logit-based distillation, first confirmed in Gemini. In early 2025, a debate arose regarding a possible distillation of DeepSeek-R1 from o1; by April 2025, it was confidently stated that DeepSeek had not done so. However, by the time of R1, it was considered more likely that some traces from o1 had been used to facilitate training, without undermining DeepSeek's innovation.

Cybersecurity: Ineffective Ban and Call for a Coherent Policy

An incident involving OpenAI and Hugging Face is cited to argue that banning access to open models would not deprive malicious actors of these capabilities. In July 2026, the emergence of autonomous hacking forms described as too cheap to measure was mentioned. In August 2026, a call was made for a national cybersecurity policy dedicated to AI, structured to observe, orient, decide, and act against emerging threats, rather than blocking models based on internal assessments of capabilities.

US–China Rivalry: Investments, Structural Advantages, and Pace

The competition over open models features distinct trajectories. In August 2025, a call was made for the United States to invest in open models for R&D purposes. Other arguments suggest that these models support research innovation and that banning open-source AI would contradict values such as education, innovation, and competition. In July 2026, vague mechanisms of federal oversight were described as conducive to confrontation or even a ban on cutting-edge open models. On the Chinese side, a history dedicated to open source and analyses of structural advantages were published in 2025–2026, along with notes from laboratories describing practices unique to the local industry. In August 2026, explanations regarding GLM-5.3 detailed how Chinese laboratories keep pace with the frontier. Technical series like Pythia and Olmo further illustrate the state of the art in fully open language models.

Strategic Milestones: Definitions, Adoption, and Data Constraints

Several texts frame the notion of open models, the reasons for their publication, and the associated risks, including a strategic perspective by Bill Gurley. In July 2024, Mark Zuckerberg outlined Meta's motivations for publishing open models. Approaches invite consideration of them as a gradient, based on parameters such as licenses, execution costs, and data access. Other analyses discuss the future economic role of open models, their likely persistent lag behind closed models, and distinct adoption dynamics. Proposals detail a safe pathway for publishing open weights. In terms of risks, a study from February 2024 found a very marginal increase in documented potential risks related to textual LLMs, while in July 2024, a significant contraction of open data was documented. Articles from 2025–2026 revisit the impact of recent Chinese models, and a summary for 2025 is available. For tracking, the ATOM Report from April 2026 compares adoption between the United States and China, a dashboard presents downloads, derivatives, and regional adoption, and a hub catalogs key models.