Reproduce “Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Moonshot AI
Moonshot AI introduces Kimi K3.1 Pro, a post-trained version of Kimi K3 and its most capable 3T-class open-weight model, with a 1M-token context window and a 2-3× faster Highspeed variant. Trained with far more RL compute across hundreds of domains, it matches or approaches leading proprietary models on coding, 3D creation, finance, and agentic benchmarks. Moonshot also removed cyber-related training data on purpose, so the open weights stay behind the frontier on offensive cyber capability. Full weights are due by Nov 4, 2026.
Key insights
- Removing cyber data works as a concrete control lever: the same pipeline without it solves 90.5% of xdev-bench-hard at pass@3, against 81.0% for the released model.
- Moonshot commits to not releasing open weights that extend the absolute frontier of cyber capability, and plans to raise the cyber capability it makes available step by step.
- On long-horizon tasks the model debugs its own setup: building a walking simulator, it probed the dynamics, found and fixed 36 errors in shoulder-muscle moment-arm calculations, and retrained the controller.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.