New Open Weight AI Models from China Renew Calls for Regulation
We’ve seen the release of two highly capable AI models from China over the past week, including Moonshot AI’s Kimi K3 model and Alibaba’s Qwen 3.8 tMax. Benchmark tests show the open models are nearly as good as top American frontier models, but at just a fraction of the cost. The new models are putting US policy in a precarious spot, as US officials seek to build and maintain a competitive advantage in the private and public sectors while American AI companies increasingly are calling for regulation of Chinese models.
Kimi K3, which Moonshot AI released on July 16, features 2.8 trillion parameters, native visual understanding, and a 1M-token context window. The company says it is the world’s first open-source model in the 3-trillion-parameter class, and is designed for frontier intelligence scenarios, including long-horizon coding, knowledge work, and agentic reasoning.
Moonshot AI says its internal suite of benchmark tests show that Kimi K3 outperforms all other AI models except for two: Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol. Kimi models, including K2, have set the upper bound of open-model sizes for nine months out of the past year, the company says in a post on its technical blog.
Kimi K3 is the biggest open-weight model (Image source Moonshot AI)
Meanwhile, on July 19, Alibaba launched Qwen 3.8 tMax, a 2.4 trillion parameter model designed for advanced coding and agentic tasks. The Chinese Web giant behind Qwen claimed tMax outperforms all other AI models except for Anthropic’s Fable 5 in terms of capability.
Calls for Regulation
The two new models demonstrate big improvement in AI performance, despite the embargo placed on China by the U.S. that prevents exports of the latest Nvidia GPUs to China. They’re also less expensive than top-tier American AI models, such as Claude Fable 5, which costs $10 per million input tokens and $50 per million output tokens. Moonshot AI charges $3 per million input tokens and $15 per million output tokens; Alibaba uses a different pricing scheme
The fact that the Chinese are developing highly capable models, and offering access to them at a fraction of the cost of frontier American models–not to mention open sourcing them to those who want to run them on their own hardware–is a major concern to US officials and American AI companies. While Chinese models are gaining traction among American companies, who are free to use any software, they are effectively banned for use by US government agencies.
Meanwhile, the top-tier AI providers increasingly are asking for regulation of AI. Dean Ball, OpenAI’s head of strategic futures and formerly a member of the Trump administration, recently said that, while Kimi K3 is a very good model, its emergence signals a need for more state action to protect American interests.
(Photo-for-Everything/Shutterstock)
“One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product, AI is a ‘public good’ which will ultimately be provided by the state as a kind of ‘digital public infrastructure,’” Ball wrote in an X post on July 17, the day Moonshot AI launched Kimi K3.
Anthropic’s CEO, Dario Amodei, has called for regulation to prevent the spread of super-capable AI models that can discover and exploit software vulnerabilities. Demis Hassabis, the CEO of Google’s DeepMind Labs, has also indicated his preference for AI engineers to be a part of the regulation discussion.
No Chinese AI in US Gov’t
The Department of Energy allows its national labs to use both proprietary and open-weight models. This includes the top American AI models from Anthropic, Google, and OpenAI, which are all closed source and therefore can’t be downloaded and customized by users. Open weight models from companies like Nvidia, Thinking Machine Labs, and Reflection AI, as well as non-profits like the Allen Institute for AI (Ai2) and others, are also OK to use. However, they don’t perform at the level of frontier models.
This creates a dilemma for AI customers who want to access the most performant AI capabilities. Anthropic, Google, and OpenAI require customers–including the US government–to pay large fees to use their models. AI companies have started to increase their fees as they seek to recoup the hundreds of billions of dollars they spent to train them.
Scientists at the national labs who are pushing the limits of AI for science and engineering as part of the US Government’s Genesis Mission project are similarly in a quandary. The government will pay for AI capability, but only up to a point. The situation is similar at NSF-funded supercomputer centers at American universities, which are actively adopting open AI models and paying for users to tap into proprietary frontier AI models.
The big questions come down to cost, capability, and security. Accessing the best AI capability costs money, but it’s unclear how much the government will pay before seeking to build its own model.
Balancing Capability, Cost, and Risk
Aron Brand, CTO of CTERA, said use of Chinese AI models shouldn’t be banned, but they should be treated as untrusted infrastructure.
“Use them for controlled, non-sensitive workloads, preferably through self-hosted deployments with strict guardrails, continuous output testing, and independent code review,” Brand tells HPCwire. “Custom fine-tuning, as demonstrated by products like Cursor that rely on retrained Chinese foundation models, can likely suppress or override many undesirable behaviors. However, we do not yet know whether it can eliminate deeply embedded capabilities or dormant triggers inherited from the base model. Until model interpretability catches up, that uncertainty remains part of the risk equation.”
Chinese models have been detected behaving differently when used in a government context, Brand added, citing a May Booz Allen study. While it’s unclear what the cause is, it’s still a cause for concern.
“Whether this behavior is deliberate or an emergent result of training data and alignment, the enterprise risk is the same: inconsistent behavior, political constraints, and the possibility of hidden backdoors or other latent behaviors,” Brand says.
There is an upside to using Chinese open-weight models, as they can deliver high performance at significantly lower costs, not to mention greater freedom in deployment.
“For startups and enterprises, they can materially improve AI economics and reduce dependence on vendors such as OpenAI and Anthropic, which are themselves navigating a whirlwind of financial pressure, regulatory scrutiny, and shifting geopolitical constraints,” Brand said. “The downside is that low cost and open weights do not equal trust.”’
This article originally appeared on HPCwire.
Related
Broadcom Announces VMware Cloud Foundation 9.1
PALO ALTO, Calif., May 5, 2026 — Broadcom Inc., a global technology leader that designs,…
NVIDIA Announces Financial Results for 1st Quarter Fiscal 2027
SANTA CLARA, Calif., May 21, 2026 — NVIDIA (NASDAQ: NVDA) has reported record revenue for…
Cerebras Systems Announces Launch of Initial Public Offering
SUNNYVALE, Calif., May 4, 2026 — Cerebras Systems, Inc. today announced that it plans to commence…
NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI
TAIPEI, Taiwan, June 1, 2026 — NVIDIA today launched NVIDIA Cosmos 3, an open world…
OpenAI Launches Deployment Company to Scale Enterprise AI Adoption
OpenAI has agreed to acquire Tomoro, giving the OpenAI Deployment Company experienced Forward Deployed Engineers…
NVIDIA Launches Nemotron 3 Nano Omni Model, Unifying Vision, Audio and Language for AI Agents
Best-in-class open omni-modal reasoning model delivers the highest efficiency and accuracy to power agentic workflows…
