Dashboard

Kimi K3.1 Pro: What Moonshot Actually Shipped

Moonshot AI announced Kimi K3.1 Pro and a faster Highspeed variant on 23 October 2026. Separating the company's own claims from the numbers circulating in coverage.

Cecilia Iona
Cecilia Iona
Senior Editor, AI & Product
25 October 20261 min read

Moonshot AI announced Kimi K3.1 Pro on 23 October 2026, along with a faster companion called Kimi K3.1 Pro Highspeed. The announcement gives it a 1 million token context window and calls it the first model in a new K3.1 series. If you build with AI, the question worth asking on a launch day is not whether a model wins. It is which claims came from the company and which came from somewhere else, because those two sets have drifted apart on this release.

What the announcement actually says

Moonshot's forum post is short. It confirms four things:

  • Two models shipped together: Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed.

  • A context window of 1 million tokens. The Kimi platform's pricing documentation puts the exact figure for the K3 series at 1,048,576 tokens.

  • Highspeed runs at two to three times the speed of K3.1 Pro.

  • Moonshot describes K3.1 Pro as its most capable 3T-class open-weight model to date, with improvements in coding, 3D creation, financial research and other agentic work.

The post includes a benchmark chart as an image. It names no benchmark, no score and no competing model in the text itself.

What only the coverage says

Three numbers are being repeated widely, and none of them appear in Moonshot's announcement text.

Claim

Where it comes from

65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra and just behind Claude Opus 5.5

Company-reported figures relayed by Unite.AI

Roughly 2.8 trillion total parameters, about 104 billion active per token

Secondary coverage, not the launch post

Full open weights by 4 November 2026

ANI and other outlets

Pricing is in the same category. Reported figures of two dollars per million input tokens and ten dollars per million output tokens are circulating, but the Kimi pricing page did not render per-token rates for K3.1 Pro when checked for this piece. It does document automatic context caching, with a default cache lifetime of five minutes and cache writes billed separately by lifetime, which matters more for your bill than the headline rate if you send repeated prefixes.

None of this means the numbers are wrong. It means they are company-reported figures passed through outlets, and a self-reported benchmark is a marketing artifact until someone else runs it. That distinction is the whole job when you read a release note, and it is the same reason release trackers disagree with each other about what shipped and when.

Why a 3T-class open-weight model matters to you

"Open weight" is not the same as open source, and the difference decides what you can legally do with the thing. Weights you can download let you run a model on infrastructure you control, pin a version that will not change under you, and keep prompts inside your own network. They do not necessarily come with training data, a permissive licence, or the right to resell access. Check the licence before you build a product on it, and read open weights versus open source if that boundary is fuzzy.

The 3T figure refers to total parameters in a mixture-of-experts design, where only a fraction activate per token. A 2.8 trillion parameter total with roughly 104 billion active is a model you cannot casually self-host. The practical access path for almost everyone reading this is the API, and the open weights are a hedge against the API changing rather than a thing you will run on a workstation.

What to do this week

Very little, and that is the correct answer most launch days.

  1. If you are not currently blocked by context length or cost, do nothing. A 1M token window solves a problem you may not have.

  2. If you are blocked, wait for the open weights and for independent benchmark runs before you plan a migration. Both are days away, not months.

  3. If you want to test it anyway, run your own task set rather than anyone's leaderboard. Our guide to evaluating a new model release before switching covers building a golden set that reflects your actual work.

  4. Write down what would have to be true for you to switch. If you cannot name it, you are reading launch posts for entertainment, which is a fine hobby but not a plan. There are calmer ways to follow AI news.

The open weights date is the thing to watch. A 3T-class model with downloadable weights changes what a small team can commit to without a vendor relationship, and that is a bigger story than any single benchmark row.

How did this land?

About the author

Cecilia Iona
Cecilia Iona

Senior Editor, AI & Product

Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.