Meta employee says Chinese open models will lose
Zengyi Qin, a member of Meta’s superintelligence lab and a core contributor to Muse Spark, has come under fire after publicly arguing that Chinese open models are unlikely to win against U.S. open models.
His case had two parts. On the technical side, Qin said Meta has an extra order of magnitude in compute and better data, and said Muse Spark would eventually surpass Chinese models such as Kimi. On the commercial side, he argued that large U.S. customers such as JPMorgan would switch to American open models for compliance reasons, leaving Chinese labs without that portion of inference revenue. He also added that Meta has Facebook and Instagram as revenue engines, while Chinese model companies rely more heavily on model-related income itself.
Developers in China quickly challenged that view. One of the most direct responses asked why Meta, which has not lacked compute or data over the past two years, still failed to hold back Chinese models. Others asked how much revenue JPMorgan had actually contributed to Kimi. Some comments went further, saying that if this reflects the reasoning level of a core Muse Spark member, it raises fresh questions about Muse’s own model performance.
The report said those objections were not simply rhetorical. Several of the claims made in the debate can be checked against public information.
Zuckerberg’s Aug. 10 essay pointed in the opposite direction
On Aug. 10, Mark Zuckerberg said in an Instagram video that Meta would release the weights for Muse Spark 1.2 and launch Muse Glimmer, an open model family designed to run on laptops. On the same day, he also published a roughly 6,500-word essay on AI.
In that essay, Zuckerberg said foreign labs currently hold several advantages because U.S. labs must comply with additional restrictions on training data. If the U.S. wants its open models to stay ahead over the long term, he wrote, policy needs to reduce that added friction.
That position clashes directly with Qin’s argument. Qin said Meta has better data. Zuckerberg said U.S. labs face more training-data restrictions, making that an advantage for foreign labs and a problem Washington should address.
The two also diverged on the commercial question. In the same essay, Zuckerberg said he does not believe restricting access to foreign open models is an effective solution. His stated goal was to make U.S. open models the best in the world by removing barriers that make them less competitive. If compliance pressure would naturally drive customers back to U.S. models on its own, there would be little reason for Meta’s CEO to spend time calling for looser policy conditions.
JPMorgan uses an in-house LLM Suite
One line from the comment section — asking how much revenue JPMorgan had contributed to Kimi — was presented as the easiest part of the dispute to verify.
Public information shows JPMorgan’s enterprise AI strategy is built around its in-house LLM Suite. The platform integrates models from multiple suppliers in a tightly controlled environment and currently connects to OpenAI and Anthropic. The bank’s public wording is that it chooses the right model for the right use case and uses open source intelligently where appropriate, but the report said no public record could be found showing adoption of Chinese open-source models.
The current regulatory direction also does not clearly support Qin’s argument. According to BlockTempo’s earlier reporting, the White House told OpenAI, Anthropic and Google in a closed-door meeting on Aug. 4 that Chinese open-weight models would be exempt from government testing under a new AI safety framework. That meant another setback for the mandatory review push backed by Anthropic CEO Dario Amodei. Nvidia CEO Jensen Huang has also said publicly that open models such as Kimi are strong and should be embraced rather than banned.
Qin has his own history in open-source AI
Qin is also a co-founder of MyShell. During his PhD studies at MIT, he worked with MyShell on OpenVoice, an open-source voice model capable of real-time voice cloning. The project has long been one of GitHub’s more popular repositories.
He has previously said that open models are central to collaboration and progress, citing Stable Diffusion and OpenVoice as examples of open AI systems that are both useful and commercially viable.
The report did not say that his current view must therefore be wrong. It said positions can change when roles change. The dispute, in that framing, is not about whether open source works. It is about whose open models will win.
Two questions raised in the report
Has Muse Spark 1.2 opened its weights?
Zuckerberg said in his Aug. 10 Instagram video that Meta would release the weights for Muse Spark 1.2 and launch Muse Glimmer, an open model family that can run on laptops. Meta’s capital expenditure this year is projected to reach as much as $145 billion.
Is JPMorgan using Chinese open-source models?
Based on public information, no adoption record could be identified. JPMorgan’s enterprise AI setup centers on its in-house LLM Suite, which integrates multiple model providers in a controlled environment and currently connects to models from OpenAI and Anthropic.