Chinese Open Models Haven't Just Won the Leaderboard — They're Now Pricing the Hyperscalers
In July, Axios reported that Chinese open-weight models — from Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai — held all five top spots on OpenRouter's weekly token-usage leaderboard. Checked directly against OpenRouter's live rankings on October 17, the roster has reshuffled but not changed character: the same five companies still occupy, or sit one slot from, the top of real inference traffic. DeepSeek's V4.1 Flash leads at 37.1 trillion tokens processed, followed by StepFun's Step 5 Preview (19.5T), Z.ai's GLM 5.3 Flash (12.9T), Xiaomi's MiMo-V2.6-Flash (11.2T), and Tencent's Hy4 preview (7.88T). The first non-Chinese model, Nvidia's free Nemotron 3 Ultra, sits at number six with 7.27T tokens — ahead of OpenAI's GPT-6 Luna (6.77T) and Anthropic's Claude Opus 5.5 (5.85T).
A gap that keeps not closing
Usage data is a lagging signal, but the benchmark gap that was supposed to justify the price differential has also narrowed. LiveBench, the long-horizon evaluation suite, placed DeepSeek in its top five for the first time this month, with leading Chinese open-weight models trailing US frontier systems by roughly 3% on score, according to Rest of World's reporting from a San Francisco panel this week. Chinese open models can still run 50% to 90% cheaper than top American closed systems on comparable tasks, and Mozilla CTO Raffi Krikorian told Axios in July that cheaper options can cost up to 50 times less for routine workloads — a gap wide enough that the three-percentage-point benchmark deficit barely registers for procurement decisions.
The clearest takeaway should be that the Chinese are very, very good at building LLMs. The point is that some of the best and most funded researchers in the U.S. may not win. Competition is hard. — Nathan Lambert, head of AI research, Trillium Labs
Western releases, same week, still short
The persistence of the gap is notable because Western labs did not sit still. Reflection shipped its first open-weight model, Mistral released an update to its large open checkpoint, and Aleph Alpha put out a new open model in the same week as the panel Rest of World covered. None changed the capability picture enough to matter, per Stanford DigiChina's Graham Webster: "If you need relatively frontier capability, right now that's Chinese open models only." Webster added that the US government itself would prefer a leading open-weight model be domestic — an implicit admission that, as of this week, it isn't.
Moonshot wants a cut of the hyperscalers, not just their downloads
The more structurally interesting move is commercial, not technical. Moonshot AI has been in talks since at least late August with Microsoft, Amazon, and Google about hosting its Kimi K3 model on Azure, AWS, and Google Cloud — and is asking for up to 30% of the resulting service revenue, according to The Decoder's sourcing of three people familiar with the discussions. No Chinese AI company has landed a comparable revenue-sharing hosting deal with a major US cloud provider before. If it closes, it inverts the standard open-weight narrative, in which giving weights away is supposed to commoditize the model layer and push value toward infrastructure and application providers. Instead, Moonshot is trying to extract rent at the distribution layer precisely because its weights are already the ones US cloud customers are pulling in volume — the leverage comes from demand the hyperscalers can't currently match with their own open checkpoints.
That demand is real and growing for reasons unrelated to geopolitics: Kristy Loke, a MATS research fellow on the same panel, noted that open weights let users "download them and use them and own them" without depending on a single vendor's API — a property with its own value to cost-conscious and sovereignty-conscious buyers regardless of where the weights were trained. One Indian university researcher cited by Axios in July estimated open-source models could eventually handle 95% of enterprise queries, leaving a thin frontier slice for OpenAI and Anthropic.
For ML engineers deciding what to fine-tune or deploy, the signal worth tracking isn't this week's benchmark table — it's that the same five labs have anchored real inference volume for three consecutive months, and one of them is now negotiating to get paid by the companies that were supposed to be disintermediating it. Washington's response so far has run through export and national-security channels rather than funding a comparable open-weight program at home; Anthropic's own Fable and Mythos models were restricted for foreign nationals earlier this year on national-security grounds, a defensive move that does nothing to address why US developers keep routing traffic to Shenzhen and Hangzhou in the first place.