I was lately welcomed to concise a collection of Congressional members and personnel on the province of open-weight models in the lens of U.S.-China competition. I’m sharing my prepared remarks as a province of the union on open models that is accessible to a broader audience.
Open tongue models are AI models anywhere their weights are publically accessible for inspection or downstream use. These are most frequently contrasted to so-called “closed” AI models. Closed models recommendation admission lone through Application Programming Interfaces (APIs) that developers can use to immediately query a model, akin GPT-4 or Claude Opus 4.5, or through products, akin ChatGPT and Claude Code.
Open tongue models chiefly are bucketed into two categories, open-weight and open-source models. Open-weight models are the most average form, specified as famous models akin Meta’s Llama, Alibaba’s Qwen, Google’s Gemma, or DeepSeek’s models. These models are governed by licenses, governing documents dictating what is allowed alongside downstream use, and are frequently accompanied by conclusion code in libraries specified as Transformers, VLLM, SGLANG, etc. Since concerning April 2025, Chinese AI companies have been the apparent chief in open-weight models.
True “open-source” models are akin to these, as they contain the weights, licenses, and conclusion code, but they additionally contain the complete data needed to reproduce the example – the training code and training data. The most notable open-source models have been built in the United States, led lately by the Allen Institute for AI’s Olmo models that I helped build in my latest 2.5 years there. The another notable open-source models are additionally built by American non-profit organizations, including OpenAthena’s Marin models and EleutherAI’s Pythia models.
Open-weight, open-source, and all another tag for a example – including closed models chiefly offered via an API – be on a spectrum. For example, Nvidia’s Nemotron models are far additional open than most open-weight models, releasing ample quantities of their training data under permissive licenses, but they’re not completely open-source since they do not publish all of the data. Closed models additionally exist on a spectrum according to what data the API reveals and the conditions of use.
We are living in a earth anywhere GLM-5.2 and Kimi K3, several of the latest, foremost Chinese models, have enacted a stage alter in the business viability of open models — crossing a akin threshold in agentic capabilities that Anthropic’s Claude Code traversed in December of 2025.
America was the first chief in open tongue models, chiefly through Meta’s Llama models, which were used extensively throughout investigation and business tasks. Chinese open-weight models surpassed American open-weight models in these two key areas concerning 18 months ago. The uncomplicated metric showing this is Hugging Face Downloads, anywhere China took the guide in July of 2025 chiefly through the achievement of Alibaba’s Qwen models. I personally keep tools to track this data, and since I archetypal published the American Truly Open Models (ATOM) Project in August of 2025, China’s download guide has grown to concerning 1.6B – alongside a total of 3.2B downloads, twice that of America’s total.
On famous capabilities benchmarks, specified as the Artificial Analysis Intelligence Index (AAII), the Chinese open-weight models have a apparent guide complete American counterparts. The top three Chinese models as of penning this on September 14, 2026 are Z.ai’s GLM-5.3 and GLM-5.3-Flash and Moonshot AI’s Kimi K3 alongside scores of 45, 42, and 44 respectively. By comparison, the foremost American models are Thinking Machines’ Inkling and Inkling Small, the two alongside a mark of 26, and Nvidia’s Nemotron 3 Ultra, alongside a mark of 23. The top American models were released in June and July of 2026, and are updated small frequently than their Chinese counterparts. For example, Chinese labs released models alongside scores complete these American models 2-6 months before the American companies got there (e.g. GLM-5 or DeepSeek V4 Pro). There is a trend of more American companies releasing models, including names akin Arcee AI, Poolside and IBM, but they are not quickly decision this achievement gap. Other benchmarks inform a akin story.
Together, Chinese open-weight models are about 2-5 months rearward the closed American frontier, alongside the open-weight American models being about 6-9 months rearward the likes of OpenAI and Anthropic. The Chinese labs are closest in tasks alongside apparent person demand, specified as agentic coding, and additional rearward on additional open-ended specialized tasks, specified as discipline or biology.
The reasons why Chinese labs can create these powerful models, notwithstanding having small resources than American counterparts, is motionless an open conversation and heavily influenced by distinct activity cultures, but is additionally influenced by a few key specialized factors. The Chinese labs publish their models faster and concentration on a slightly narrower allocation of tasks, flattering them slightly on community benchmarks. Releasing faster helps them mark higher since all the labs are making accordant progress, so formerly you “finish” a example to be released, it is a snapshot of achievement at that stated period — labs anywhere that period is afterward lean to mark higher. Still, the models built by the Chinese labs are genuinely powerful and portray genuine competition to the American industry. This competition volition not decrease meaningfully as the closed labs place vulnerabilities in their API offerings which allow distillation.
Distillation is most impactful in new domains and does not create it trivial to create a universally powerful final model. I evaluation that if distillation was completely prevented, e.g. alongside know-your-customer (KYC) tools at Anthropic and OpenAI, the gap from the strongest American models to Chinese open-weight models would lone addition by 1-2 months.
For example, the Chinese labs are quickly changing their posture towards paying for training data in 2026. Earlier in the year, the top Chinese labs including Moonshot AI and Z.ai had a powerful penchant towards construction data workflows in-house, but by the summer they had begun to buy the cutting border data – challenging RL environments for agentic tasks – from the two established American companies and new Chinese startups.
With the advancement of open importance models in China towards the frontier of capabilities, and the latest records of expanding risks about frontier models in areas specified as cybersecurity (e.g. the OpenAI-HuggingFace incident), there’s expanding regulatory doubt on how continued releases can allow a safer ecosystem?
A structural difficulty in open-weight models is that there are few productive methods for stopping pieces of open application from reaching bad actors. If an attempt was made to restrict admission to the strongest open-weight models from China since they amplify risks, the parties who would be set rear are American businesses. We have an example of this – HuggingFace used a Chinese open-weight example to comprehend the cyberattack since closed models would not answer their requests. Thus, managing the risks of open-weight models often comes downward to ecosystem preparation.
Open-weight models are becoming an essential tool for AI diffusion, and the finest way to get onward of these risks and unbalanced relationships anywhere American companies depend on models built in China is to continue to allow funding in open models in the US. Ownership of open models allows improved coordination and preparedness of risks that are earth in their nature during accelerating diffusion of AI services throughout the family economy.
Open-weight tongue models have grown substantially in broad involvement and financial viability in 2026, allowing first glimpses of additional straightforward ways to difference acceptance of models from the US, China, or elsewhere on top of Hugging Face metrics. One example is OpenRouter usage. OpenRouter is a famous LLM conclusion phase that supplies a sole interface to toggle between models, open and closed, from the US and China. This phase is chiefly known for trying distinct open-weight models. The phase has shared use data for the top models since Jan. 1, 2025, and shown growth in use from ~1T tokens processed from open models in a week of September 2025 to ~80T tokens per week today. In that time, Chinese models have grown from ~70% market portion to complete 80% of usage. Other platforms that are designed to commercialize open models display akin data, specified as the open-source coding delegate OpenCode, which shows an inference quantity of ~95% or higher alongside Chinese models.
These open platforms are the finest approximation of open example use we have – a ample proportion of open example use is on platforms that do not disclose per-model breakdowns, specified as Together AI or Fireworks AI, and in personal deployments for endeavor applications.
Many notable innovation companies and startups have been construction on Chinese open-weight models for their AI features, specified as Harvey, the lawful agent, Cursor, the coding agent, and DoorDash’s use of Kimi models, Airbnb’s use of Qwen, or Perplexity’s use of DeepSeek. These notable companies are the tip of the iceberg, anywhere a ample swath of younger Silicon Valley startups are construction on Chinese models in command to have low-cost, elastic options. There is a growing trend of American startups and companies entering endeavor agreements alongside Chinese example labs in command to get approval to use their models in their products – a new form of cross-border innovation collaboration I have not witnessed in my career.
The basis of innovation on Chinese models extends additional into the AI ecosystem. To a archetypal command approximation, most of scholarly investigation is conducted on Alibaba’s Qwen family of models. Having met multiple members of the Qwen guidance squad during my trip to China, they are extremely invested in and intentional concerning this category of adoption, which volition not be uncomplicated to claw rear to American models.
To quantify the acceptance of open models throughout academia, I scanned all document in the 5 most famous ML categories of arXiv (cs.AI, cs.CL, cs.CV, cs.LG, stat.ML), the preprint phase famous in AI research. The results plainly track my understanding of the evolving guidance in AI research, showing LLMs becoming a foundational tier of ML investigation – mentions of any open example were 2% in January of 2023 and 50% in September of 2026 – and the foremost function change from the U.S. to China in the identical period period.
For example, in April to May of 2023, a few months following Meta’s first Llama (a backronym, Large Language Model Meta AI, archetypal released in Feb. of 2023), concerning 2,600 of 12,000 new AI/ML document on arXiv mentioned at smallest one notable open example family. Of all those scanned papers, ~5.5% mentioned Llama and ~1% mentioned a Chinese model. In the autumn of 2024, during Llama’s peak, concerning 23% of document mentioned Llama alongside concerning 7.5% mentioning Qwen, the most straightforward Chinese competition. Today, Llama has misplaced its guide in academia, being mentioned in concerning 21% of document still, which is notable longevity, but Qwen’s portion has risen to 30% of papers. Overall, any Chinese open importance example is mentioned in complete 40% of papers, complete the U.S.’s 30%, alongside China’s portion continuing to grow.
This shows that we plainly have a lot of activity to do in command to re-establish the U.S. as the residence of AI investigation in the era of open-weight tongue models. There are signs of hope.
In our research, we discover that American models of comparable capabilities-to-size regions to their Chinese counterparts get adopted at disproportionate rates. In the final twelvemonth we’ve seen OpenAI’s archetypal open-weight models since ChatGPT, gpt-oss, rotate into among the most adopted open-weight models of all time. Since then, Google’s Gemma 4 models have been several of the lone ones always to display akin acceptance numbers to Qwen’s most famous small models, and Nvidia’s Nemotron models have humble acceptance notwithstanding many additional capable models at the identical size point.
The narrative of open models in 2026 is one of establishing financial relevance. This is the convergence of many stories throughout the AI ecosystem, summarized as:
The capabilities gap from open to closed models accessible to users has been decreasing complete the final 3 years. This varies by task, but can be estimated as a 2-5 duration gap in capabilities. With capabilities general progressing so fast, this has seen open-weight AI models unlock significant markets in 2026 and points to additional inflection points in the near future.
Open example use is exploding in high-value industries (e.g. application engineering, legal services, financial services), indicating an increase of an substitute ecosystem to the finest closed models. Platforms offering conclusion chiefly on open models, from Together, OpenRouter, Fireworks, Baseten, etc., are seeing amazing growth as the archetypal winners of an open example post-training economics (other layers contain finetuning APIs specified as Thinking Machines’ Tinker). This is blended alongside many anecdotes from specialized personnel in the AI industry that uses open-weight models specified as GLM-5.3 as an substitute to Claude or GPT because of a blend of speed, lesser prices, customizable offerings, and privacy.
Chinese AI companies are the apparent leaders in open importance models. Relative to 2025, anywhere Chinese models akin DeepSeek R1 shook the AI earth alongside surprise, the American AI labs have been recovering in their positions alongside open-weight models, but notwithstanding additional significant funding in the US, the Chinese labs regularly are producing notably stronger models adored by many types of users.
Distillation of American AI models by Chinese labs does not explain the complete narrative of their success. Distillation is an industry norm method of training another AI example on the outputs from a normally stronger model. The method is most prevalent in the Chinese AI industry, which has used essential exploits to extract reasoning traces and additional data from American companies’ products that are not completely secured. The finest estimates are that distillation helps decrease the achievement gap of Chinese companies related to the American frontier by 1-2 months.
Chinese models, particularly Alibaba’s Qwen family, are established as a foundational tier of investigation and betterment throughout academia and local example users. In latest months, Chinese open importance models were mentioned in 38% of AI papers, complete the U.S.’s 28% – and the Chinese portion is expanding much faster than its American counterparts. This, alongside alongside another governmental factors and the closed nature of foremost American AI companies, is contributing to an accelerated decrease in America’s guide as the preeminent AI investigation hub in the world.
Open importance models are entering the capability levels anywhere new risks, e.g. cybersecurity, can be enabled by many open-weight models being available, necessitating an ecosystem flat reply in preparation. This new era of risks is additionally enabling a duration of governmental uncertainty, anywhere there is regulatory notice on the strongest AI models, but enormous doubt on how guideline would be legally enacted. At the identical time, many researchers and engineers depend on open models because of additional permissive safeguards, anywhere the closed models specified as Claude and GPT frequently refuse crucial cybersecurity protective activity or existence discipline research.
For additional data, perspective the Interconnects Dashboard.
In 2026 the Chinese labs are plainly maintaining their position as the leaders of the open-weight AI ecosystem. This comes as open-weight models have passed an inflection item in financial viability and in the visage of risen action from American labs as example competition. The foremost Chinese labs do not appear to be meaningfully challenged, as they develop their endeavor and investigation acceptance globally.
This landscape of open models comes at a crucial period in the broader AI ecosystem. We’re seeing OpenAI and Anthropic obtain enormous steps onward alongside their latest community models, and at the identical period call for coordinated attention on how we oversee the next phase of AI progress. What is happening in the confines of a few AI labs today, particularly alongside extreme endowment and compute density, is a precursor to what volition shortly appear in the open example ecosystem. Open models are going to be the substrate for everyone alternatively in the earth exterior of the few true frontier AI labs, to harness an acceleration in application engineering and another computational practices. This represents a significant origin of gentle power, influence, and possible for the organizations that allow this broad admission to transformative intelligence.
With this forthcoming coming soon, we need to collectively remain humble concerning the exact way open models volition take. There are a lot of unknowns alongside open models – e.g. we don’t have fine data on how they’re used in countries another than the U.S. and China. With the allocation of ML training ability being broad, ie tens of organizations and thousands of group that are inside a twelvemonth of the frontier of capabilities, it is a matter of when, not if, open models cross the achievement thresholds that allow new workflows. The company method have to be to comprehend how to use this broadly accessible, open intellect for fine during proactively mitigating the possible harms.
Thank you to Florian Brand and Kevin Xu for feedback and/or suggestions for this work. For additional investigation notifying this post, see the open-source AI study list.
