← Back to feed

The two largest AI powers are moving in opposite directions on open-source model policy — and the collision is reshaping how frontier weights will flow globally.

In Washington, the Trump administration and the AI industry have been discussing a capability framework for US open-source models. The framework would set a capability ceiling for freely releasable model weights based on the current capabilities of leading Chinese open-source models. The mechanism is straightforward in concept: if a US lab builds a model that does not exceed what China’s best open-weight models can already do, releasing the weights is acceptable. If it exceeds that threshold, different rules would apply.

In Shanghai, Xi Jinping addressed the World Artificial Intelligence Conference and said China is ready to be more open about artificial intelligence. The statement came as Chinese labs continue to release frontier open-weight models — DeepSeek V4 Pro, GLM-5.2, Kimi K3, Qwen3.6-Plus — that are now competing directly with the top proprietary models on global intelligence indices.

The Strategic Structure

The US capability framework, as reported by The Washington Post, has a built-in paradox. Chinese labs are releasing increasingly powerful open-weight models. Each new release from DeepSeek or Moonshot or Alibaba raises the empirical capability ceiling of “what China can do in open-source.” If US policy pegs the allowable open-source tier to Chinese capabilities, then China’s release cadence effectively sets US open-source policy.

China holds all ten spots on the current Open-Weights Intelligence Index. The frontier gap between Chinese and US closed models has compressed from 13 points to 6 points on the Artificial Analysis Intelligence Index over the past year. Industry sources expect that Chinese labs will eventually release Mythos-class model weights — the capability tier currently restricted to vetted government and research partners in the US. When that happens, the US capability cap would either auto-adjust upward or become de facto inoperative.

What the Framework Would Accomplish

From a US policy standpoint, the framework addresses a specific concern: US-origin open-weight models trained on US compute, IP, and research could accelerate adversary AI development if released without restrictions. The proposed threshold approach attempts to thread the needle — allowing US companies to participate in the open-source ecosystem without releasing capabilities that materially exceed what is already globally available.

From the industry standpoint, the framework is better than a blanket export control on model weights, which several labs and researchers have warned would drive development offshore. Anthropic, in a 2028-focused policy paper, called for locking down compute now; the capability framework takes a different angle, focusing on release thresholds rather than training restrictions.

The AI industry expects Chinese Mythos-class model weights to become available for free download at some point. That expectation is part of the policy calculus: if the US imposes strict open-source caps and China releases Mythos-tier weights anyway, US labs would be at a competitive disadvantage in the developer ecosystem without having achieved any security benefit.

Enforcement Is the Missing Piece

Any capability framework depends on a definition of “capability.” Benchmark scores are the obvious metric, but as interconnects.ai has noted, enforcement of model-level capability thresholds has no working precedent. Intelligence Index scores can be measured. But a 1.6-trillion-parameter open-weight model with MIT license, once released, cannot be un-released. The lag between a US lab producing a frontier-class model and that model being assessed against a Chinese open-weight benchmark could span months.

The framework also faces a definitional problem from the Chinese side. Xi’s statement at WAIC positions China as pro-openness. Beijing has simultaneously moved to lock advanced Chinese AI models inside China’s borders for domestic use. The public posture of openness and the domestic policy of access restriction are not contradictory from Beijing’s perspective — but they complicate any bilateral framework that relies on Chinese open-weight releases as the reference benchmark.

The Current Open-Weight Landscape

The backdrop makes the policy stakes concrete. China’s GLM-5.2 (Z.ai) leads open-weights on the Intelligence Index at 51. Kimi K3 (Moonshot AI) is 4th globally across all models — open and proprietary — at 78.5 on LiveBench. DeepSeek V4 Pro ships under MIT license at $0.87/M output. Qwen3.6-Plus leads on several agentic coding benchmarks. The open-weight frontier is already competitive with proprietary models from Western labs on multiple benchmark dimensions.

In that context, a US capability cap tied to Chinese open-source capability is not a mechanism to prevent capability parity — that has already happened on several benchmarks. It is a mechanism to slow the rate at which future US-origin capabilities enter the global open-weight commons. Whether that distinction has the strategic value its architects intend will depend entirely on whether China continues releasing increasingly capable open-weight models — which, based on the past eighteen months, it will.