Kimi K3.1 Pro: A Large AI Model Anyone Can Download, With Its Hacking Skill Cut on Purpose

Kimi K3.1 Pro is a 2.8-trillion-parameter AI model from Moonshot AI that anyone can download and run. Before release, Moonshot removed hacking-related training data: on its own test the model solved 81.0% of problems, against 90.5% without the removal.
artificial-intelligence
cyber-security
Author

Kabui, Charles

Published

2026-11-09

Keywords

kimi-k3-1-pro, moonshot-ai, open-weights, mixture-of-experts, cyber-capabilities

Download as PDF

Kimi K3.1 Pro is a large AI model from the Chinese lab Moonshot AI. Anyone can download its weights (the 2.8 trillion numbers it learned) and run it on their own servers, which suits companies and researchers with big hardware. The files total 1.56 terabytes, about three 512 GB laptop drives. It is a mixture of experts model: only 104 billion of those numbers work on each word, like a hospital where each patient sees a few specialists. Moonshot announced it on October 23 and published the weights about November 4.

Artificial Analysis, an independent testing site, ranks it highest among open-weights models. At launch it gave Kimi 55 and fourth place. On October 27 it moved to index version 5, built mostly on private tests, and the score became 46, 14 index points behind the leader. Each test task takes Kimi 37.2 minutes, against a 13.0-minute median for similar open models. A company that sells access to the model needs a separate deal with Moonshot once its revenue passes $20 million over 12 months.

Before release, Moonshot removed cyber-related data from its training to hold back its hacking skill. Moonshot’s blog does not say whether fine-tuning, which the licence allows, could restore it. On Moonshot’s own test, xdev-bench-hard, the model gets three tries per recently disclosed software flaw to write a working attack. A version without the removal solved 90.5% of problems; the final model solved 81.0%, and Notebookcheck reported that no outside group had verified either figure.

Read More: Deltafin, a research tool that runs the earlier 2.8-trillion-parameter Kimi K3 on one workstation and Anthropic’s test of another open-weights model, GLM-5.3, whose refusal filters can be removed

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Kimi {K3.1} {Pro:} {A} {Large} {AI} {Model} {Anyone} {Can}
    {Download,} {With} {Its} {Hacking} {Skill} {Cut} on {Purpose}},
  date = {2026-11-09},
  url = {https://toknow.ai/posts/kimi-k3-1-pro-open-weights-hacking-skill-cut/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Kimi K3.1 Pro: A Large AI Model Anyone Can Download, With Its Hacking Skill Cut on Purpose.” https://toknow.ai/posts/kimi-k3-1-pro-open-weights-hacking-skill-cut/.