24 OCT 2026 — Moonshot AI, the Chinese company behind the Kimi models, released Kimi K3.1 Pro on 24 October and says it deliberately trained the model to be worse at hacking than it could have been. The model is live now in Kimi's apps and API. Its weights, the part that makes it open, are promised by 4 November.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, whose Claude Opus 5.5 model is one of the systems Moonshot compares K3.1 Pro against, and whose Claude Code tool Moonshot used to run several of its tests. Readers should weigh that interest.
What Moonshot released
In its launch post, Moonshot calls K3.1 Pro "the first model in our new K3.1 series and our most capable 3T-class open-weight model to date". It has a context window of one million tokens. A second version, K3.1 Pro Highspeed, is said to run two to three times faster.
The new model is not a fresh build. Moonshot describes it as "a post-trained version of Kimi K3", trained with much more compute in the post-training stage. Kimi K3, according to Moonshot's July post, is a 2.8-trillion-parameter model, introduced on 16 July with its weights promised by 27 July. A public Kimi-K3 repository now sits on Moonshot's Hugging Face page. When we checked on 24 October, no K3.1 repository was listed there.
Both K3.1 models can be used through the Kimi website and apps, the Kimi Work desktop app, the Kimi Code tool and Moonshot's API. For companies on Kimi Enterprise, access to K3.1 Pro is off by default and an administrator has to switch it on.
The benchmark claims
Moonshot supplied all the scores below, and no outside lab has reproduced them. On Terminal-Bench 4.0, a test of coding agents working in a terminal, the company reports 65.7% for K3.1 Pro. That puts it ahead of OpenAI's GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. Moonshot's footnotes say it ran its own model through the Claude Code harness for that test and took the rival scores from published blogs or the official leaderboard.
On BenchCAD, which asks a model to rebuild 3D objects by writing CAD code, K3.1 Pro scored 92.7% with tools. Claude Opus 5.5 scored 96.2% and GPT-6 Astra 95.9%, so the open model still trails there. On an internal long-horizon coding test, Kimi Code Bench 3.0, the score rose to 61.2% from K3's 55.95%.
The company also reports a sharp fall in made-up facts. On its own stress test, K3 got 23.4% of individual claims wrong. K3.1 Pro got 8.3% wrong. Moonshot adds its own warning: the test uses prompts "deliberately crafted to probe situations where models are prone to errors", and the rates "do not represent overall error rates in daily use".
Training it to hack less
The most unusual section of the post covers cyber capability. Moonshot says vulnerability-research skill rose steadily as post-training scaled, and it traced the rise to security material sitting inside its general training data and among its reinforcement-learning tasks. For the open-weight release, it says it removed cyber-related data from both stages while keeping ordinary coding tasks.
It measured the effect on xdev-bench-hard, an in-house set of exploit tasks built on recently disclosed bugs in software such as the Linux kernel and Node.js. A checkpoint trained without the filtering solved 90.5% of problems. The released model solved 81.0%. Moonshot says that is still a big improvement on K3, because better general coding carries over to security work.
Moonshot frames this as a release rule. "We aim not to release open weights that would extend the absolute frontier of cyber capability," the post says. It also says it intends "to progressively raise the cyber capability available in open weights", arguing that teams who run a capable model on their own machines can audit code that rarely gets a security review. Moonshot's claim that K3.1 Pro sits behind proprietary models released at least three months ago rests on its own testing. None of these results have been checked by an outside lab.
Cheaper than the model it replaces
On the Kimi API pricing page, kimi-k3.1-pro costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20. That undercuts K3, listed at $3 and $15. The Highspeed version costs twice as much as K3.1 Pro, at $4 and $20. The page does not state the currency. All K3 series models share a context window of 1,048,576 tokens.
The same page tells K2 users to move. From 31 October 2026, new users will no longer be able to use K2 series models, and Moonshot plans to retire the whole series by the end of December. For anyone comparing bills, our LLM cost calculator works through input and output rates separately.
What to watch before 4 November
For now, the open-weight label is only a promise. Until the files appear, K3.1 Pro can only be used through Moonshot's own services, and none of its benchmark figures can be tested by anyone outside the company on their own hardware. Moonshot named a July release date for K3, and a public repository exists today, although we could not confirm the exact day it went up.
Moonshot is also expanding its sales effort. RECATOOLS reported on 23 October that Moonshot is hiring in Singapore to sell its models. India's ANI news agency, in a report carried by The Tribune, described the launch from the company's statement.
View author profile → · Editorial policy
