The numbers that define the Chinese position

Forbes (April 2026) documented Qwen's numbers with a precision that deserves attention: 1 billion cumulative downloads on Hugging Face by March 2026, reaching this milestone faster than any other open-source model family in history. In February 2026 alone, 153.6 million downloads — more than Meta, DeepSeek, OpenAI, Mistral, NVIDIA, Zhipu, Moonshot, and MiniMax combined in the same period. Qwen captured more than 50% of all global open-source model downloads by March 2026. There are 180,000+ models derived from Qwen on Hugging Face — more than Google and Meta combined.

DeepSeek V4, launched on April 29, 2026, is described by the Council on Foreign Relations (April 2026) as "the most significant update since the release that shook global markets." Open, multimodal — capable of generating text, images, and video in a unified framework — and explicitly designed to run without the most advanced NVIDIA chips, which are subject to American export controls.

Remote OpenClaw (July 2026) documents the state of the benchmarks: Chinese labs now hold four of the top five positions in open-weight AI, with GLM-5 (Zhipu AI), Qwen3.5 (Alibaba), Kimi K2.5 (Moonshot AI), and DeepSeek V4 leading in different capability dimensions. The best Chinese line still lags approximately 9 points behind the best proprietary models from OpenAI, Anthropic, and Google — but the gap closed much faster than most industry forecasts estimated.

The strategy the US did not anticipate: efficiency, not scale

The analysis by the US-China Economic and Security Review Commission (March 2026) captures the most important dynamic: Chinese labs did not try to compete directly with the American scale paradigm. Instead, they developed an efficiency strategy — "do more with less" — that proved especially valuable in a context of chip export controls.

DeepSeek R1, launched in January 2025, was the most visible sign: a model that achieved near-frontier performance with a fraction of the computational resources of equivalent American models. Abhishek Gautam's analysis (March 2026) documents the cost equation: Qwen 3 costs US$ 0.38 per million tokens, compared to US$ 15 per million for Claude 3.5 Sonnet or US$ 10 for GPT-4o. "Chinese models deliver 75–85% of GPT-4o's quality at 10–15% of the cost."

For developers and startups with budget constraints, this equation is transformative. Researchers at Stanford and UC Berkeley trained high-performance models using Qwen with costs as low as US$ 30–50. This democratizes access to AI research in a way that American frontier models, at US$ 100 million per training, simply cannot.

The two-loop strategy: openness as a geopolitical advantage

The Neuron (July 2026) analyzes what it calls a "phase shift" in Chinese strategy: "For a year, Chinese AI labs have gained global mindshare by making capable models cheap, useful, and widely available. This openness has been useful for China. It spread Chinese models to startups, research labs, cloud marketplaces, and enterprise experiments worldwide. It also pressured American labs on price and release cadence."

The US-China Commission (March 2026) articulates the "two-loop strategy": open models spread technological influence globally (outer loop) while strengthening the domestic AI ecosystem (inner loop). It is a platform strategy — by making models so accessible that startups and researchers around the world use them as a base, China builds technological dependence that goes beyond individual products.

The real limits: chips, content restrictions, and trust

An honest analysis of the Chinese position includes the real limits. Remote OpenClaw (July 2026) documents one that rarely appears in benchmark comparisons: all Chinese models — DeepSeek, Qwen, GLM, Kimi — have hard-coded content restrictions on topics sensitive to the Chinese government. Independent testing confirms that these models refuse or provide answers aligned with government positions on Taiwan, the Tiananmen Square protests, and Xinjiang. In some cases, models actively insert messages aligned with the Chinese government rather than simply refusing. For use in corporate and government applications in democracies, this is a real obstacle.

The chip problem is also real but is being partially bypassed. American export controls block access to NVIDIA's most advanced chips — H100, A100, and their successors. But the Chinese efficiency strategy was partly a response to this restriction: if you can't have the fastest chips, develop algorithms that need less compute. DeepSeek proved this is possible up to a certain point.

Digitalinasia.com (May 2026) notes that this also explains China's growing interest in developing domestic AI chips independent of NVIDIA — a long-term bet that has not yet delivered competitive frontier chips but is being accelerated as a matter of technological sovereignty.

What this means for the global ecosystem

CFR (April 2026) captures the most important strategic implication of DeepSeek V4: "China's latest model trails American competitors in benchmarks. But it may not need to win the performance race to reshape the geopolitics of artificial intelligence."

For developers and organizations outside the US and China, the proliferation of high-quality open-source models — both American (Meta Llama) and Chinese (DeepSeek, Qwen) — creates a third way that didn't exist in 2023: access to world-class AI capabilities without dependence on any of the large proprietary platforms. This has particularly significant implications for Brazil and Latin America, where the cost of frontier APIs is amplified by currency disparity.

Abhishek Gautam (March 2026) summarizes the state: "For developers, this is straightforward good news. More competition means lower prices, faster capability improvement, and more choices. The geopolitical dimensions are real and should inform your vendor risk assessment."