Kimi's weights are due Wednesday, and Google's No. 1 model has no release date
2 Nov 2026 · News
Moonshot says anyone will be able to download its best model by Wednesday. Google's top model is going to vetted cyber defenders first, with no date for everyone else.
Moonshot AI, the Beijing lab behind Kimi, promised to publish the full weights of Kimi K3.1 Pro by 4 November. Those weights would make it the best model anyone can download and run. The model that tops the main leaderboard, Google's Gemini 4 Carbon, is only going to a vetted group of cyber defenders for now. As of Monday morning, neither has changed. We think Moonshot delivers this week and Carbon stays gated through the end of the year.
What Moonshot promised
Moonshot launched Kimi K3.1 Pro in late October. Its launch post says, in two places, that the full model weights will be released by 4 November. Anyone can already use the model through Kimi's apps and API. The weights are the bigger deal. With them, anyone can run, test or modify the model on their own machines.
They weren't out on Monday morning. Moonshot's Hugging Face page had no K3.1 Pro repository, and its newest model, Kimi K3, was last updated on 1 September. Moonshot also hasn't said which license K3.1 Pro will use. Kimi K3 shipped under Moonshot's own custom license.
Moonshot thought about cyber risk too, and came to a different answer than Google. It says it removed cyber-related training data from the open release. On one hard hacking benchmark, a version trained with that data passed 90.5% of tasks and the final model passed 81.0%. Moonshot says it aims "not to release open weights that would extend the absolute frontier of cyber capability." It also argues that "making K3.1 Pro available in a non-discriminatory way benefits people who need dual-use capabilities for defensive purposes."
What the American leaders hold back
Google's Carbon announcement on 28 October says Carbon ranks first on the Artificial Analysis Intelligence Index at 59. Access starts with "a subset of cyber defenders," then more "trusted partners" through Google's Fairwind Program. Google says it is taking part in the U.S. government's voluntary pre-release testing and will open Carbon to developers, businesses and consumers "as soon as possible." As of Monday, Carbon wasn't in the Gemini API release notes, and Artificial Analysis still lists it as "not yet publicly available."
OpenAI is rationing in a different way. As we reported on 29 October, OpenAI picked the people who get to check its physics, and that committee chose to hold back the proof. OpenAI's own release post admits that choosing the first reviewers "creates asymmetry of information."
The index shows why Moonshot's promise matters. In the new version of the index, K3.1 Pro scores 46 and is the top open-weights model, with a note that its weights aren't released yet. The best models you can actually download today, from DeepSeek and Zhipu's GLM, score 32. That's 27 points behind Carbon and below Anthropic's small Claude Haiku 5.5. Moonshot's release would close more than half of that gap in one day.
Merchants and mercantilists
Venkatesh Rao's essay Cognitive Mercantilism, which we used on Sunday to look at who's actually paying for AI, has a line that fits this week. Mercantilism, he writes, "was not a philosophy of, by, or for merchants." It belonged to statesmen and aristocrats. He says the same instinct runs through AI. Many elites, he writes, treat intelligence as "a scarce, dangerous resource that must be managed by credentialed custodians, kept expensive, used directly or indirectly for conquest and treasure, and denied to rivals." AI leaders, he adds, act like mercantilists "in direct proportion to their desire to be more than mere merchants."
The American labs do sell their models, by the token. Buyers rent access, and the labs decide who gets the top model and when. Moonshot is promising to hand over the goods themselves. So the lab acting most like a merchant is the one in Beijing.
There's a practical reason to care. About 60% of the new index runs on private or held-out tests that outsiders can't see. Open weights are the one release that lets anyone check a lab's claims, including Moonshot's claims about its own cyber limits.
Our call
Moonshot's Kimi K3.1 Pro weights will be public on its Hugging Face page by 7 November 2026, and Gemini 4 Carbon will still not be generally available to developers on 31 December 2026. We give Moonshot three days past its own deadline to allow for time zones and small delays.
The Carbon half is the riskier one. Google's previous model, Gemini 4 Argon, also went to cyber defenders first and reached the Gemini API 13 days later, on 13 October. Carbon is different. It ranks first on the index, Google says it's in government pre-release testing, and it gave no date. We're wrong if the K3.1 Pro weights aren't on Hugging Face by 7 November, or if developers can call Carbon through the Gemini API before the year ends.