Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, and the numbers are worth a close look. The model is a mixture-of-experts system with 2.8 trillion total parameters, 104 billion active per token and a 1M-token context window, according to Moonshot's own launch post. It is available now through Moonshot's API. The full model weights are promised by November 4, 2026, and as of the launch coverage they were not yet downloadable.
What happened and when
Moonshot announced K3.1 Pro on October 23 and positioned it as its most capable open-weight model. SiliconANGLE and The Next Web both reported the launch and the November 4 weights timeline the same day. Simon Willison recorded US API pricing of $2.00 per million input tokens and $10 per million output tokens as of that date. Moonshot also offers a higher-speed tier that it says runs two to three times faster at twice the price.
Moonshot's stated strengths are full-project coding, including games built in Unreal, Unity and Godot, parametric 3D modeling in Blender and CAD, and multimodal input. Those are vendor claims. We have not run them.
The cyber trade-off
Moonshot says it removed cyberattack-related material from training before the open release, so the model would not push the limits of what freely available models can do, according to The Next Web. SiliconANGLE reports the same scrub targeted the post-training dataset. Moonshot's own exploit test shows the gap it created: a version trained without the cuts solved 90.5 percent of the test, the released model solved 81 percent, and Moonshot's earlier K3 solved 33.3 percent. Those figures come from Moonshot and have not been independently checked.
What it means for solo operators and marketers
A 1M-token window changes how research work gets built. A year of client call transcripts, a full campaign history or a sprawling content library can go into one prompt instead of being chopped into pieces and stitched back together. That is useful for an AI solopreneur who would rather spend an afternoon on judgment than on retrieval plumbing.
Open weights matter for a different reason. If the November 4 date holds, teams that need client data to stay inside their own infrastructure get a serious option. At this size, though, self-hosting is a job for a well-funded shop, not a weekend project. For most of us, the API at $2 and $10 per million tokens is the practical path, and pricing is only current as of October 23, so check it again before you budget around it. The license terms were not in the coverage we reviewed, so read them before anything touches client work.
Our take
Twenty-five years in digital marketing has taught me to distrust any benchmark a vendor publishes about itself, and Moonshot's exploit numbers are exactly that kind of benchmark. Still, the instinct behind the scrub is interesting. I co-founded Urchin, and the early analytics debates were never about whether the software could be gamed. They were about whether the people relying on it could trust what it reported. A lab choosing to trim one dangerous capability before the weights leave its hands is a precedent worth watching, and a far more interesting story than another leaderboard flex.
Our advice is to get an API key this week, run one real workload at the published rates, and put November 4 on your calendar. Wait for the license text before anything near sensitive client data goes in. The model may well be as good as Moonshot says. The only number we would trust without a second look is the one on your own invoice.