Kimi K3.1 Pro: a 3T-class model with open weights coming in November

On October 23, China’s Moonshot AI released Kimi K3.1 Pro, the first model in the new K3.1 series, together with a faster variant, Kimi K3.1 Pro Highspeed. According to the blog post, it is Moonshot’s strongest 3T-class model (roughly 3 trillion parameters), it has a 1M-token context window, and the full weights are due to be published by November 4, 2026.

What’s new in Kimi K3.1 Pro?

This isn’t an entirely new model. It’s Kimi K3 after a much longer post-training run. Moonshot improved, among other things, its RL (reinforcement learning) framework, long-context handling and the environments where the model learns agentic tasks. The gains are supposed to show up in coding, 3D modeling and financial analysis.

The most interesting points from the blog post:

1. The Highspeed variant has the same capabilities but runs 2-3 times faster.
2. The model works in thinking mode only, with three reasoning_effort levels: low, high and max (the default is max). It also accepts images and video.
3. On Kimi Hallu Bench Hard, an internal hallucination test, the share of false claims dropped from 23.4% to 8.3% (with tools). Moonshot says this is a stress test, not the everyday error rate.
4. On Kimi Code Bench 3.0, an internal benchmark of long coding tasks, the score rose from 55.95% to 61.2%.
5. kimi.ai and the Kimi app gained a 3D Make feature: the model creates 3D models you can edit online and download for 3D printing.

Interestingly, Moonshot also showed a longer agentic demo. Using the open body model human-atlas (296 skeletal elements and 402 muscles), the model built a gait simulator over 167 tasks and 420 reviews. The simulated figure went from falling over after a few seconds to 11.2 s of continuous walking, and along the way the agent found and fixed 36 bugs in the shoulder muscle model.

Most of these numbers are Moonshot’s own measurements, so let’s take them with a grain of salt. There is already an independent result, though: on the Artificial Analysis Intelligence Index, Kimi K3.1 Pro scores 55 points.

Why is everyone talking about it?

First, open weights at this scale. A 3T-class model with downloadable weights is rare, even if for now (until November 4) Artificial Analysis lists it as closed.

Second, a decision about cybersecurity, which is what most of the discussion on Hacker News is about. For the open-weight release, Moonshot removed cybersecurity-related data from mid-training and from the RL environments, keeping ordinary coding tasks. On xdev-bench-hard, an internal test of finding and exploiting vulnerabilities in code, the version without these changes solved 90.5% of tasks, while the final K3.1 Pro solved 81%. Moonshot says it doesn’t want to publish open weights that push the frontier of offensive cyber capabilities much further, but it plans to raise that bar gradually to help with auditing legacy code.

Some commenters see this as a sensible compromise; others point out that once the weights are out, that data can simply be fine-tuned back in.

How to try Kimi K3.1 Pro?

The model is available in several places:

1. Kimi.ai and the Kimi mobile apps (iOS, Android, HarmonyOS).
2. Kimi Work, the desktop app, from version 3.1.0 (Windows and Macs with Apple silicon).
3. Kimi Code, where you pick the model with the /model command. According to the Kimi Code documentation, K3.1 Pro requires the Plus plan or higher (1M context from the Pro plan), while Highspeed requires the Pro plan and uses roughly 2 times more of your quota.
4. API with the identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed. According to the API guide, it is an OpenAI Chat Completions-compatible interface, and access requires topping up your account with at least 1 USD.

According to Artificial Analysis, the Kimi API price is 2 USD per million input tokens and 10 USD per million output tokens.

Once the weights are online, K3.1 Pro will join the models you can run yourself, although at this scale it’s a job for a server room rather than a laptop (more on hardware in the guide how to run AI locally). Until then, it’s another candidate to plug into an agent harness, much like MiniMax M3.1 Flash.

FAQ

Does Kimi K3.1 Pro have open weights?

Not yet. Moonshot plans to publish the full weights by November 4, 2026. The open-weight version will have cybersecurity-related data removed.

How much does Kimi K3.1 Pro cost?

According to Artificial Analysis, on the Kimi API it costs 2 USD per million input tokens and 10 USD per million output tokens. In Kimi Code, the model is available from the Plus plan.

How is Kimi K3.1 Pro Highspeed different?

It has the same capabilities as K3.1 Pro, but according to Moonshot it runs 2-3 times faster. In Kimi Code it requires the Pro plan and uses about 2 times more of your quota.