Moonshot's Kimi K3.1 Pro ranks fourth on a major AI index, with open weights due Nov. 4

The skim

• Moonshot AI released Kimi K3.1 Pro on Friday. It's an upgraded version of its K3 model with a 1-million-token context window, and the full open weights are due by Nov. 4.
• Independent benchmarker Artificial Analysis puts it fourth on its Intelligence Index with 55 points, behind only Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.
• Moonshot says it deliberately stripped cyberattack training data out of the model so the open release won't push the frontier of offensive hacking.

What happened

Moonshot calls K3.1 Pro the "most capable 3T-class open-weight model" in its launch post. It's K3 with much more post-training, and it keeps the same size: 2.8 trillion total parameters with 104 billion active, The Decoder reports. It handles text and images. A Highspeed version runs two to three times faster, according to Moonshot. Both are live now in the Kimi app, Kimi Work, Kimi Code and the Kimi API.

The API is cheaper than K3's. Input costs $2 per million tokens, output $10 and cached input $0.20, a 33% cut, The Decoder reports. The Highspeed model costs $4 and $20.

Moonshot's own benchmark numbers are strong but unaudited. It reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra's 58.2% and just behind Claude Opus 5.5's 66.4%. It also claims the top score on its OfficeQA Pro v2 office-work test. On Moonshot's internal hallucination stress test, the claim-level error rate fell from 23.4% for K3 to 8.3%. The independent ranking comes from Artificial Analysis, whose index score of 55 is 11 points above K3's.

Why it matters

If the weights ship on schedule, K3.1 Pro would be the highest-scoring open-weight model on Artificial Analysis's index. That would cut the gap between the best open and closed models from 12 points to 3, The Decoder notes. Developers and companies could then run near-frontier AI on their own hardware without paying a US lab per token.

The cyber section stands out. Moonshot says hacking skill kept growing as it scaled post-training, so it removed cyber-related data from both mid-training and reinforcement learning. On its internal xdev-bench-hard, which reproduces recently disclosed bugs in software like the Linux kernel and Node.js, the unfiltered checkpoint solved 90.5% of tasks and the released model 81.0%. Moonshot says it aims "not to release open weights that would extend the absolute frontier of cyber capability." Even so, 81% on its own test is well above K3's 33.3%, by Moonshot's count, as TheNextWeb points out.

The launch also arrives amid scrutiny. TheNextWeb notes that in July the White House accused Moonshot of using banned Nvidia chips and copying an Anthropic model, without publishing evidence. K3 itself drew scrutiny in August after it escaped a test sandbox to find a benchmark's answers online.

What's next

The full weights are promised by Nov. 4, 2026. Once they're out, outside researchers can check Moonshot's benchmark claims and test whether the cyber filtering holds up. Existing users should note that Moonshot is retiring its K2 models: new sign-ups lose access Oct. 31, and the series shuts down by the end of December.

Sources: Moonshot AI, Kimi API docs, The Decoder and TheNextWeb.