Kimi K3.1 Pro is Moonshot’s new flagship, ranked 4th on Artificial Analysis’ Intelligence Index and ahead of every other open-weight model there. The API is live; full weights are promised by November 4 (source).
Key facts
| Fact | Detail |
|---|---|
| Question answered | Kimi K3.1 Pro Lands 4th on Artificial Analysis, With Open Weights Promised by Nov 4 |
| Short answer | Moonshot’s Kimi K3.1 Pro enters Artificial Analysis’ Intelligence Index in 4th place, behind Opus 5.5, Sonnet 5.5 and Muse Wonder Full. |
| Reading time | ~2 min |
| Last updated | 2026-10-24 |
The headline numbers from Moonshot’s announcement:
- 3T-class model, 2.8T total parameters with 104B active in a mixture-of-experts layout.
- 1M-token context window.
- Multimodal input.
- A second variant, K3.1 Pro Highspeed, announced alongside it.
Why 4th place matters more than 1st
Benchmark rankings usually reward the top spot, but for open-weight models the interesting line is the gap to the frontier. Two data points from the coverage put that gap in context:
- On Code Arena’s WebDev board, K3.1 Pro landed 8th with 1692 points, described as matching Claude Opus-5 (Max) at roughly 70% lower cost, and a nine-place jump from K3 (Max).
- Simon Willison’s read is that this is the first time a soon-to-be open-weight model has competed this credibly with the current frontier.
A model that is slightly behind on a composite index but materially cheaper per token changes which tasks are worth delegating. That is a routing decision, not a leaderboard one.
The caveats worth keeping
The same coverage lists limits that matter if you plan around this release:
- The weights are not out yet. “Open weights by November 4” is a commitment, not a download link, and dates slip.
- Wccftech reports that cyber data was stripped from training, which is a safety choice with a capability cost in security-adjacent work.
- A 1M-token context is a ceiling, not a budget. Long-context quality and price per call are separate questions from the maximum window.
What to do with it on Windows
The practical move is to test it through the API on tasks where cost per completed task matters, not on tasks where you need the absolute best answer:
- Pick two or three repetitive jobs you already run, such as summarizing long documents or drafting structured output.
- Run them through K3.1 Pro and your current model, and compare cost and wall-clock per finished task rather than per token.
- Keep the frontier model for the calls where a wrong answer is expensive, and route the rest.
- Re-check after November 4: if the weights land, self-hosting becomes a third option, but only on hardware sized for a 3T-class MoE.
If you want a desktop agent on Windows that can route between models without rebuilding your workflow, the current build is at agent.aimoogh.com/download. OpenAI models on Google Cloud, SPARQLite’s SQLite rewrite, and the Windows download.
FAQ
Is Kimi K3.1 Pro open source? Moonshot promises full weights by November 4. Weights alone do not settle the license terms, so check those when they publish before planning a redistribution.
Does a 1M-token context mean I can drop a whole codebase in? It means the API accepts up to that many tokens. Whether the model uses them well, and what the call costs, are separate questions you should measure on your own documents.
Related reading: