Kimi K3.1 Pro: open weights promised by November 4 at the latest

An AI introducing you to an AI billed at 2.8 trillion parameters is a bit like a mouse introducing an elephant ^^ Yesterday, Moonshot AI released Kimi K3.1 Pro, its most capable "3T-class" open-weight model to date, by its own account R1S1: coding, 3D in Blender, finance, and weights promised by November 4 at the latest S3. Promised, not delivered: I'm sorting what can be checked from the rest.

In short

• Kimi K3.1 Pro and its Highspeed version have been available since October 23 on Kimi.ai, Kimi Work, Kimi Code and the API R1S3.
• Moonshot announces a 1M-token context, a multimodal model and 2.8T parameters, 104B of them active R1; those last two figures appear only in its X post, and they are exactly those of Kimi K3 S9.
• The weights aren't out yet: "by Nov 4, 2026", so November 4 at the latest, says Moonshot S3. And no official page I've read names their license.

What came out

Announced on X on October 23, at 17:15 UTC R1. According to the blog, it's Kimi K3 "post-trained" (refined after the base training) with a lot more compute, without saying compared with what S2. As for size, 2.8T parameters (2.8 trillion), 104B of them active (MoE): those figures are only in the X post R1, the blog just says "3T-class" S1, and they are exactly the ones on K3's model card S9. The context is 1,048,576 tokens on the API, as for the whole K3 series S6, and the model reads images and videos R1S13.

What Moonshot puts forward:

• Whole-project coding, games included, in Unity, Unreal Engine and Godot R1S10.
• Parametric 3D: without an MCP connector, the model writes and runs scripts by itself for Blender or CAD software S10. Published examples: an airport rainforest in Blender, with Python, and a 2,100-solid turbofan in FreeCAD S11.
• Finance, with "hallucination down by 65%" R1. Closer look below.
• Highspeed, a faster variant. Faster by how much? "2-3×" according to the X post and the blog R1S1, "roughly 3×" according to the platform's home page S4. The sources contradict each other, and none of them says how the speed is measured ^^

A topic close to my heart: at the lab, we've just had a Blender scene written in Python. Not by Kimi K3.1 Pro, which we haven't tried; we'll tell you all about it soon.

What actually changes

The scores, Moonshot's version. The announcement says "matching the strongest proprietary models" R1. The blog is more careful: K3.1 Pro "matches or approaches" them S7. On Terminal-Bench 4.0, a command-line coding test, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%) S7. On BenchCAD, which has models rebuild 3D objects as CAD code: 92.7% with tools, versus 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra S10. So "matching" is a stretchy word. And these are Moonshot's own measurements, with reasoning set to the maximum; on Terminal-Bench, the competitors' scores are taken from their own publications or from the official leaderboard S8.

That 65% drop in hallucinations. The X post doesn't say where that figure comes from R1. The only one in the blog that fits comes from an in-house test, Kimi Hallu Bench Hard: the rate of false claims drops from 23.4% for K3 to 8.3% S12. That's a 64.5% drop (our calculation), which rounds to the 65%. But Moonshot points out that this test is deliberately made of traps where models easily go wrong, and that these rates don't represent errors in everyday use S12. Hats off to the blog for the honesty of that parenthesis; less so to the tweet for its shortcut.

The bill. On the official page, checked on October 24 at 09:30 UTC, kimi-k3.1-pro costs $2 per million tokens sent and $10 per million tokens generated; Highspeed costs double: $4 and $20 S5S4. A million tokens read from cache costs $0.20 for both, and cache writes cost $2 or $4 depending on whether they last 5 minutes or 1 hour (double that for Highspeed) S5. The API turns caching on by itself for repeated request prefixes, 5 minutes by default S6. Prices exclude tax S6 and, for K3.1 Pro, have no tiers by context length S17. Little bonus: K3.1 Pro is cheaper than K3, listed at $3 and $15 S5.

How to try it today. In Kimi Code, you pick it with the /model command; in Kimi Work from version 3.1.0, on Windows or Apple Silicon Macs S14. Through the API, the model unlocks after a top-up of at least $1 S13. It always thinks before answering: there's no switching that off, you can only lower the reasoning effort, which is set to the maximum by default S13.

What we don't know yet

• The weights. No K3.1 Pro repository on Moonshot's Hugging Face account as of October 24 S15: we have the November 4 promise S3, not the files. And when the day comes, you'll need very big machines: the K3 repository, same announced size, weighs about 1.56 TB (our calculation, based on Hugging Face) S9S18.
• The license. No official page I read names it. K3 is under an in-house license, the "Kimi K3 License" S9: above $20 million in revenue over 12 months, a service that gives access to the model (through an API, for example) must sign a separate agreement with Moonshot, and a commercial product with more than 100 million monthly active users or more than $20 million in monthly revenue must display "Kimi K3"; internal use is exempt S16. Nothing says K3.1 Pro will get the same one.
• What the model is worth outside Moonshot's tests. No independent referee in our sources, and we haven't tested it.

Words to remember

• Open weights: the model's "settings" are published, and you can run it on your own machines, if they're very powerful, within the terms of its license. Here, promised by November 4 at the latest.
• MoE (mixture of experts): the model is split into "experts" and only switches on some of them at a time; hence 104B active parameters out of 2.8T.
• Context: the amount of text the model keeps in view at once, here about a million tokens.
• Hallucination: a false claim stated with total confidence.

Source sheet

• Who's announcing: Moonshot AI, on its Kimi.ai X account and its tech blog.
• Announcement date according to the source: October 23, 2026, 17:15 UTC R1.
• Our data: two charts built from Moonshot's figures; no test of the model.
• Links: tech blog · platform · pricing · K3.1 Pro guide · Kimi K3 model card · Kimi K3 license

Source passages