Key Definitions
Before reviewing the support matrix, it helps to distinguish how these three deployment models differ in practice:
- Closed System (Proprietary): The model weights, code, and training data are private. Access is provided strictly via hosted APIs or platforms (e.g., ChatGPT, Claude API).
- Open Weights: The trained parameters (weights) are released for download so you can run, host, or fine-tune the model locally. However, complete training datasets, data filtering pipelines, or full training code are generally withheld, and usage may be governed by custom community/commercial licenses.
- True Open Source: Complete transparency — model weights, source code, and training recipes/data formats are fully released under OSI-approved permissive open-source licenses (e.g., Apache 2.0, MIT).
Top Company Support Matrix
| Company | Closed System | Open Weights | Open Source (OSI Approved) | Notable Models / Key Notes |
| Meta (Muse Spark, Llama ) | No | Yes | Partial | Flagship Llama series released as open weights under Meta’s Community License (not strictly OSI due to commercial usage caps). |
| DeepSeek | Yes | Yes | Yes | Offers hosted APIs (DeepSeek-V3/R1), but releases weights, architecture, and code under permissive MIT licenses. |
| Zhipu AI (GLM) | Yes | Yes | Yes | Operates hosted platforms while releasing open-weight models (GLM-4 / GLM-5 series) under permissive MIT licenses. |
| MiniMax | Yes | Yes | Yes | Operates commercial APIs while publishing open-weight architecture releases (MiniMax-M1 / M3 series) under permissive open licenses. |
| Mistral AI | Yes | Yes | Yes | Hybrid strategy: proprietary APIs (Mistral Medium) alongside permissive Apache 2.0 / MIT models (Mistral 7B, Mixtral). |
| Moonshot AI (Kimi) | Yes | Yes | Yes | Provides API platforms (Kimi) while releasing open-weight architectures (Kimi K-series) under permissive open licenses. |
| SpaceX.AI (Grok) | Yes | Yes | Partial | Hosted frontier models (Grok-3/4), but released open weights for Grok-1 and open-sourced developer tools like Grok Build. |
| OpenAI (ChatGPT) | Yes | Yes | No | Primary focus on closed systems (GPT-4o, o1), with selective open-weight releases (GPT-OSS-120B/20B, Whisper). |
| Anthropic (Claude Fable,Mythos,Opus,Sonnet) | Yes | No | No | Strictly closed-system API offerings (Claude 3.5 / Opus). Rejects open-weights releases due to safety concerns. |
| Google (Gemini) | Yes | Yes | Partial | Closed frontier models (Gemini) alongside an open-weight model family (Gemma) under custom terms. |
| Alibaba (Qwen) | Yes | Yes | Yes | Offers API services while releasing the Qwen series as open weights/open source under Apache 2.0. |
Industry Trends
- Hybrid Approach: Most companies maintain closed, commercial API platforms for their absolute flagship models while offering smaller or specialized open-weight variants for developer ecosystem growth.
- Strictly Closed: Anthropic remains one of the few top-tier frontier developers maintaining an exclusively closed system model.
Most Companies’ Stance
Open Letter to Support
Dozens of tech giants, startups, and research institutions jointly signed an open letter on June 24, strongly urging U.S. policymakers not to suppress “open-weighted” AI models.
This open letter, released on Friday, comes as the White House weighs whether to ban Chinese open-source hardware on national security grounds. The letter is unusually united by almost all of the major U.S. tech companies, including Nvidia, Microsoft, Meta, Alphabet, OpenAI, A16z, Dell, Palantir, Cisco, IBM, GitHub, and the Linux Foundation, among 32 other organizations.
Only Anthropic has yet to express its support.
Reasons for Corporate Support
Microsoft and other signatories stated in the letter that Open Weights broadens access to the AI economy, noting that organizations can build on advanced models without having to train them from scratch or pay for them.
The tech giant added that Open Weights fosters competition, ensuring that the benefits of AI are widely shared, rather than concentrated in the hands of a few players.
Open Weights allows every organization to match the right model to the right task at the right cost, reserving frontier-scale capabilities for truly cutting-edge problems and enabling efficient, specialized models to run everywhere else.
Most American companies would benefit
80% of American AI startups use open-source AI models from China. The Little Tech Association, comprised of nearly 200 Silicon Valley startups, has urged the Trump administration not to restrict China’s open-source AI technology.
The White House’s consideration of banning Chinese AI models from the US market has triggered widespread panic in Silicon Valley’s startup community. Nearly 200 Silicon Valley companies, including the well-known startup accelerator Y Combinator, jointly wrote to President Trump on July 22, warning that a complete ban would cause hundreds of American startups to “die instantly,” while a few AI giants like Anthropic would reap the benefits.
According to a report by Politico cited by Guancha.cn, this letter from the newly formed Little Tech Association was also copied to Michael Kratsios, Director of the White House Office of Science and Technology Policy, and is dated July 23, 2026.
Advantages of Open Weights
Technical Advantages
Unlike closed AI systems, open weights models allow developers and enterprises to download pre-trained parameters (called weights) and run them on their infrastructure. Essentially, this concept is similar to the open-source software movement that revolutionized computing decades ago.
Unlike fully open-source AI, open weights models typically provide pre-trained model weights but don’t necessarily release the underlying training data or source code.
Three heads are better than one
Cybersecurity defenders need advanced AI to combat increasingly sophisticated attacks. Open weights models allow more researchers to identify vulnerabilities, improve security measures, and independently test systems, rather than relying solely on the original developers.
Reduced Costs, Making It Easier to Use
Training advanced AI models can cost billions of dollars. Open weights models address this challenge by allowing developers to build on existing AI models instead of creating them from scratch. This lowers the barrier to entry for smaller companies in the AI field.
Promoting Competition
Open weighted models not only encourage competition among AI developers, but also among cloud service providers, chip manufacturers, software companies, and AI application developers.
“This competition stimulates innovation, reduces costs, and broadly extends the benefits of AI to our economy,” because open-source AI lowers the cost of building AI products.
Opponents’ Arguments
The US is inclined to ban open weighted models
OpenAI and Anthropic have recently been actively lobbying the US government, warning of the potential risks posed by Chinese open-source AI models. According to five sources familiar with the matter, they specifically mentioned the potential threat posed by Chinese open-source AI models to the US. US Treasury Secretary Scott Bessent and President Trump’s science and technology advisor, Michael Kratsios, have also participated in discussions about protecting US AI technology.
Protecting Vested Interests
Creators of closed AI models like OpenAI, Anthropic, and Alphabet face clear economic incentives to maintain exclusive control over their top technologies.
The US Fears Loss of Control
Former Google CEO Eric Schmidt admitted that China’s open-source AI is no longer under US control, and the gap has narrowed from the previously estimated one to two years to only about six months. He acknowledged that US control over Chinese chip hardware is “beginning to fail,” and the gap between Chinese AI models and top US technology has narrowed from the previously estimated one to two years to only about six months, which is almost negligible in the development of AI. What worries him even more is that China’s full embrace of the open-source path in AI means it is “no longer under US control.”
Controversy surrounding distillation
Competitors Steal by Distillation
A major concern surrounding open-source AI is model distillation, where smaller models are trained using the output of larger models. Recent high-profile cases include allegations that Chinese developers Alibaba and Moonshot AI distilled Anthropic’s Claude model and used it to build their own systems.
Western companies have long been using distillation
It’s not just Chinese manufacturers that use distillation; this technology has long been widely adopted by many Western technology companies. Currently, companies opposing American companies’ use of Chinese AI models and accusing Chinese AI companies of using distillation to steal their research results include OpenAI, Anthropic, and Alphabet—the three beneficiaries—as well as Mistral, SpaceX AI, Meta, and Microsoft, who are also using distillation to enhance and improve their own AI models.
Western Also Distill Chinese AI Companies
In fact, many Western AI companies distill Chinese models during research and training. In 2025, Mistral was exposed for distilling DeepSeek models, plunging into a public relations crisis: the technical community discovered that some of its models were highly similar to DeepSeek in their generation style, and a former Mistral employee revealed that the company deliberately concealed the distillation process, misleading the results into being presented as self-developed technology.
In April 2026, Meta also stated that Muse Spark’s training used several third-party open-source models, including Qwen from Chinese tech giant Alibaba, as well as models from OpenAI and Google. However, the practice of using Chinese models clearly goes against the stance of some US policymakers and high-level technologists. In response, a Meta spokesperson stated, “Like other companies in the industry, Meta uses techniques such as filtering and refining to learn from publicly available AI models under strict safeguards to improve our own models.”
Thinking Machines’ first model, Inkling, primarily uses DeepSeek-V3 in its hybrid expert architecture. Its cold start training also utilized data from open models such as K2.5 from the Chinese AI startup Moonshot AI.
Cursor’s Composer 2 coding model is built directly on Moonshot AI’s Kimi K2.5 through a licensing partnership, and overlaid with Cursor’s own training data. Cursor acknowledges that its Composer 2 coding model is built on Moonshot AI’s Kimi K2.5 through a licensing partnership, and then overlaid with its own training data.
Chinese Cost 60-90% Less Than US
Chiese is cheaper and very close to top US models
On Arena.ai’s front-end capability rankings, KimiK3 ranked first, and GLM-5.2 ranked fourth, surpassing the previous world’s strongest models, Fable 5 and GPT 5.6 Sol. The “Zhipu” model’s ranking as number one globally shocked Silicon Valley, with Zhipu’s stock price surging over 40% intraday on June 22, 2026.
The president of OpenAI stated that Kimi K3 is “quite good,” believing that Chinese models may only lag behind the US by four months. Anthropic indicated that the US may only have about six to nine months left to lead China in AI models.
May 2026, for the first time in history! Alibaba’s Qwen 3.7-Max model has broken into the top four of the global programming rankings, with the top three being Anthropic’s Claude models.
Cost Difference and Current Market Price Advantage
Reasons of the Cost Gap
Top US AI labs (such as Anthropic, which spends huge sums monthly on rented computing power) face enormous infrastructure and R&D expenditures; in contrast, Chinese models often employ open weight strategies and benefit from infrastructure support, significantly reducing the price of end-user APIs.
Related articles
- “OpenAI, the Generative Artificial Intelligence rising star and ChatGPT“
- “The king of AI Alphabet (Google) “
- “The pros and cons of investing in Alphabet amid AI evolution“
- “DeepSeek routed the global AI and stock“
- “Chinese AI progress and top companies“
- “The Importance of Small Language Models (SLM) in AI, how is Apple AI different from others?“
- “China ditch US AI chips and decides to go its own way“
- “AI inference chips vs. training chips“
- “Comparison of AI chips GPU and ASIC“
Disclaimer
- The content of this site is the author’s personal opinions and is for reference only. I am not responsible for the correctness, opinions, and immediacy of the content and information of the article. Readers must make their own judgments.
- I shall not be liable for any damages or other legal liabilities for the direct or indirect losses caused by the readers’ direct or indirect reliance on and reference to the information on this site, or all the responsibilities arising therefrom, as a result of any investment behavior.
