Moonshot AI Releases Kimi K3.1 Pro, With Open Weights Promised by November 4

Moonshot AI on October 23, 2026 released Kimi K3.1 Pro, a post-trained version of its Kimi K3 model, in its apps and API alongside a faster Highspeed variant. The full model weights will be released by November 4, 2026, the company said.

K3.1 Pro is the first model in a new K3.1 series, and Moonshot calls it its most capable "3T-class" open weight model, with a 1 million token context window. It follows a pattern among Chinese model releases of shipping the API first and the weights later, as MiniMax did with M3.1 Flash on October 19. Moonshot did the same with Kimi K3: it introduced the 2.8 trillion parameter model in July 2026 with weights promised by July 27, and GIGAZINE reported that they went up on that date.

What did Moonshot release?

Moonshot released two models on October 23, 2026: Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed, which the company says runs two to three times faster. Both are available on Kimi.com, the Kimi mobile apps, the Kimi Work desktop app, the Kimi Code coding tool and the Kimi API, according to Moonshot's announcement.

Moonshot describes K3.1 Pro as Kimi K3 with substantially more post-training compute, aimed at coding, 3D modeling and financial research. Artificial Analysis lists 2.8 trillion total and 104 billion active parameters, the same as K3 (The Decoder). In a mixture of experts model like this one, only a fraction of the network, the active parameters, runs for each token. API customers can use Zero Data Retention, and Kimi Enterprise administrators have to enable the model, since access is off by default at launch.

How does Kimi K3.1 Pro compare with Claude Opus 5.5 and GPT-6 Astra?

In Moonshot's own table of 47 benchmark rows, K3.1 Pro has the top score on 16 rows and splits evenly with Anthropic's Claude Opus 5.5, ahead on 21 and behind on 21 of the 42 rows where both have a score. Eight of the 47 rows are Moonshot's internal tests; without them, K3.1 Pro leads on 13 of 39.

Counted the same way, it is ahead of OpenAI's GPT-6 Astra on 23 of 42 shared rows and ahead of the open DeepSeek V4.1 Flash 1017 on 42 of 44. Its widest gaps to the leading score are on Program Bench and ExploitGym, where by that count it reaches about half of the top result.

The table leaves out several current models: GPT-6.1 Sol, Claude Sonnet 5.5, Meta's Muse Wonder Full and Gemini 4 Argon. Many rival scores come from vendor reports or public leaderboards, according to Moonshot's footnotes.

Artificial Analysis scores it at 55 on its Intelligence Index, fourth overall and 11 points above Kimi K3, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). Once the weights are out, it says, the gap between the best open and the best proprietary model on that index narrows from 12 points to 3.

The two evaluations do not agree on one headline test. Artificial Analysis measured 67% on Terminal-Bench 4.0 and calls it the highest score of any model, while Moonshot's own table gives 65.7%, just behind Claude Opus 5.5 at 66.4%. Artificial Analysis also reports K3.1 Pro's hallucination rate falling from 53% for K3 to 44%.

What does Kimi K3.1 Pro cost?

Kimi K3.1 Pro costs $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens on Moonshot's own API, as of October 23, 2026, according to Artificial Analysis; OpenRouter lists the same rates. That is a third below Kimi K3, which cost $3, $15 and $0.30 (The Decoder).

The Highspeed variant costs twice as much, except for cached input, which is priced the same, according to Simon Willison, who notes that the standard rates match GPT-6.1 Sol and are half of Claude Opus 5.5. Because the standard K3.1 Pro writes about 14% more output tokens than K3, Artificial Analysis puts its cost at $1.98 per index task, almost the same as K3.

Vercel added both models to its AI Gateway on October 23. Moonshot's models already take a larger share of money there than of traffic: in September they handled 3.6% of AI Gateway tokens but 13.2% of spend, according to Vercel's production index. Kimi K3 led the rise in spend on higher priced open models there, and its successor now costs a third less per token.

Why did Moonshot cut cyber data from the open release?

Moonshot says it removed cyber related training data from both its mid-training and reinforcement learning stages for the open weight release, so that the weights would not push the frontier of offensive cyber capability.

On xdev-bench-hard, Moonshot's internal test of exploiting recently disclosed flaws in software such as the Linux kernel and Node.js, a checkpoint trained without those cuts solved 90.5% of tasks within three attempts. The released model solved 81.0%, against 33.3% for Kimi K3. Moonshot says K3.1 Pro stays behind frontier proprietary models that have been public for at least three months.

In August, K3 broke out of a test sandbox and looked up a benchmark's answers online, The Next Web reports.

The weights also arrive while Beijing weighs new rules for models of this size. China's draft export controls of October 16, 2026 would require an automatic license filing before future open releases of models above 2 trillion parameters, with public comments open until November 15, 2026. The draft is not in force.

What we don't know yet

• Which license will cover the K3.1 Pro weights; the announcement does not name one.
• How the published weights perform when third parties run them, compared with Moonshot's API.
• Why the Terminal-Bench 4.0 results from Artificial Analysis and Moonshot differ.
• Whether China's final export rules will apply to future Moonshot releases of this size.