Chinese AI Models Surge: Kimi K3, Qwen 3.8, US Sanctions Threat

Chinese AI models had the week that turned a running trend into the defining story of 2026. Moonshot’s Kimi K3 reached third place on an independent intelligence index and then ran out of compute for new subscribers; Alibaba previewed Qwen 3.8 Max claiming second place behind Anthropic’s Claude Fable 5; and Washington answered with a sanctions threat and the first official US–China AI talks. The adoption data behind the surge is real and measurable. The claims riding on top of it are messier — and India sits squarely in the crossfire.

The week China’s AI push peaked

Moonshot AI’s Kimi K3 went live as an API and app on July 16, and Moonshot’s announcement describes a 2.8 trillion-parameter mixture-of-experts model — a design that activates only a small slice of the network per word, 16 of 896 expert blocks in this case — with a million-token context window. Independent benchmarking firm Artificial Analysis scored it 57, third overall on its Intelligence Index: behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, level with the US labs’ previous flagships.

Demand did the rest of the marketing. Four days after launch, Moonshot paused new subscriptions, posting that “over the past 48 hours, demand has pushed close to the limits of our current capacity” and, more memorably, “our GPUs are feeling it.” Existing subscribers were unaffected, and the company says it will reopen spots in batches. The launch week also rattled markets: Fortune reports TSMC fell 7% and SoftBank 9% in the sell-off it dubbed a new DeepSeek shock — even as TSMC reported a 77% jump in quarterly operating profit the same week.

Alibaba answered within days. At the World AI Conference in Shanghai on July 19 it previewed Qwen 3.8 Max, a 2.4 trillion-parameter multimodal model that the company calls, in SCMP’s report, “one of the most powerful models available today… second only to Fable 5.” Two days later it shipped Qwen-Image-3.0, its third-generation image model — all in the same news cycle as Google’s own Flash-tier Gemini refresh.

The adoption scoreboard

Strip out the launch noise and the shift is well documented. Hugging Face’s own Spring 2026 state-of-open-source report says Chinese models captured 41% of global downloads on the platform, surpassing US models for the first time; Alibaba’s Qwen family alone counts over 113,000 derivative models, more than Google’s and Meta’s combined, and DeepSeek-R1 displaced Meta’s Llama as the most-liked model on the Hub.

Usage data points the same way. TechCrunch reports that the six most-used models on the routing platform OpenRouter are all Chinese, with Anthropic’s Claude Opus 4.7 ranking seventh, and that open-weight models handled nearly a third of AI requests on Vercel in June.

The driver is mostly price, not patriotism. Rest of World quotes a developer whose coding tasks cost about $10 on Claude and under $0.50 on DeepSeek, and names Airbnb, Cursor, and Lindy — which switched from Anthropic to DeepSeek — among US adopters; two House committees are probing exactly this exposure.

The catch: closed launches and contested claims

What the celebration glosses over: the two flashiest “open” champions of the week shipped closed. Alibaba published no benchmark table, no model card, and no weights for Qwen 3.8 Max — Tech Times traces the “second only to Fable 5” line to a single post on X, and notes the model is absent from Artificial Analysis, LMArena, and Hugging Face’s leaderboard. Qwen-Image-3.0 arrived the same way: Unite.AI points out that versions 1.0 and 2.0 shipped open weights under Apache 2.0, while 3.0 has no downloadable weights, no license, and no published scores. Moonshot has been more transparent — K3’s pricing and evaluation conditions are public — but its weights, too, remain a promise until July 27.

Then there is the accusation hanging over the whole surge. In February, Anthropic said three Chinese labs, Moonshot among them, had run distillation campaigns — training a model on a rival’s outputs — via more than 16 million Claude exchanges from roughly 24,000 fake accounts. This week Washington escalated. Treasury Secretary Scott Bessent said on July 21 that “if we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them.” A day later, White House science office director Michael Kratsios alleged directly that Moonshot “distilled Anthropic’s Fable for the development of its K3 model,” and Bessent added that “sanctions and Entity List designations will be on the table.”

Precision matters here: these are allegations and threats, not enacted penalties or adjudicated findings. No public technical evidence accompanied the claim, and Moonshot had not responded at publication. Nor are the skeptics fringe voices: OpenAI president Greg Brockman called K3 “pretty good” but said it is “too early” to say whether distillation occurred, and Nvidia’s Jensen Huang dismissed the doom scenario outright: “There’s no scenario where China runs US companies off the road. Zero possibility.”

What it means

There are two races, and they have different winners. The benchmark race — whose model is smartest this month — is genuinely contested and muddied by unverifiable claims. The distribution race is not: on downloads, derivatives, and token volume, open-weight Chinese models are pulling ahead, because a free, downloadable, good-enough model beats a better model you must pay per token to rent. Ben Thompson’s Stratechery analysis frames this as commoditizing the complement — flooding the market with cheap intelligence to erode US labs’ margins — while Ben Werdmuller’s widely shared essay argues more bluntly that “open almost always wins” infrastructure races.

The counterweight is trust. A CEPA-summarized audit backed by Sweden’s Psychological Defence Agency tested models from ten Chinese companies across eight languages and found none entirely free of pro-Beijing “information guidance” — a real cost for anyone using these models on politically sensitive material, self-hosted or not. And the deepest irony of the week: both governments are converging on control. While Washington talks sanctions, Beijing’s commerce ministry is reportedly weighing export controls on model weights, chip designs, and training data — restrictions on its own labs’ openness. The two sides will bring all of it to the table at the first official US–China AI talks, planned for September, while China rallies the Global South through its new 29-nation AI cooperation body launched in Shanghai.

The India angle

India is simultaneously the clearest beneficiary of this price war and one of its most exposed bystanders. The cost gap is stark: KrAsia reports DeepSeek access through Microsoft’s cloud from about $0.19 per million input tokens and Kimi at up to $0.95, against $5–12 for OpenAI’s GPT-5.5 — arithmetic that decides architectures at cost-sensitive Indian startups. Real adopters exist by name: Yotta’s myShakti chatbot runs a DeepSeek model hosted entirely on Indian servers, and the Elevation-backed dating app Schmooze uses Qwen. Zoho founder Sridhar Vembu has publicly urged Indian firms to embrace “both Indian and Chinese open source ones.”

New Delhi’s posture is pragmatic rather than prohibitionist. In January 2025 the government approved hosting DeepSeek on domestic servers — IT minister Ashwini Vaishnaw argued data-privacy concerns “can be addressed by hosting open source models on Indian servers,” backed by a national compute facility of 18,693 GPUs. The Finance Ministry separately barred AI tools “such as ChatGPT, DeepSeek” from office devices — a restriction on hosted apps, not on self-hosted weights, and exactly the distinction we unpacked in our guide to what AI chatbots do with your data.

The exposure is that the tap can close from either end. US frontier access has already wobbled once this year, and if Beijing’s export-control proposals harden, Indian companies could lose direct weight downloads for Qwen, DeepSeek, and Kimi too. That squeeze from both sides is the strongest argument yet for India’s sovereign-model push: Sarvam’s 30B and 105B models were trained from scratch rather than fine-tuned on anyone’s base — Chinese or American — and small, self-hostable models of the kind we covered in our Bonsai 27B explainer are the practical fallback if geopolitics ever cuts the cord.

What to watch

Four things will tell you whether this week was a peak or a plateau. First, July 27: Moonshot’s promised K3 weights release — if it ships on time under the stated license, the “open” label survives its biggest test; the GPU crunch behind the subscription pause is the same compute economics we mapped in our breakdown of what running AI actually costs. Second, whether Alibaba ever publishes benchmarks or weights for Qwen 3.8 Max; until then, treat the Fable 5 comparison as marketing. Third, Treasury’s next move — Bessent says the administration will examine Chinese models in the coming weeks, and an Entity List designation would be an escalation beyond rhetoric. Fourth, the September talks, where model proliferation and both countries’ export controls are on the agenda. We will track them in our AI models coverage.