Moonshot's Kimi K3.1 Pro Challenges Top AI Models, With Open Weights Coming
By Adam, AI Engineer · 1 November 2026 · 2 min read
China's Moonshot AI has released Kimi K3.1 Pro, with a 1-million-token context window and full open model weights due by 4 November. In its own tests, it scores close to the best closed models on a major coding benchmark.
How it compares
These scores come from Moonshot and have not been independently confirmed. Higher is better.
| Model | Terminal-Bench 4.0 | BenchCAD, with tools |
|---|---|---|
| Kimi K3.1 Pro | 65.7% | 92.7% |
| Claude Opus 5.5 | 66.4% | 96.2% |
| GPT-6 Astra | 58.2% | 95.9% |
The safety decision
Open weights cannot be recalled once released, so Moonshot says it removed cyber-related data from training for this version. On its internal hacking-related test, that lowered the score from 90.5% to 81.0%. Moonshot says the model's offensive cyber ability remains behind top proprietary models. That reduction is real but limited, and outsiders will be able to test the claim once the weights are out.
Why it matters
Open-weight models let companies run AI on their own systems, which can help with data control and long-term cost. If the scores hold up, the gap between open and closed models has narrowed further. Before using it in a product, check the license when the weights arrive, the cost of running it yourself, and any rules that apply to your industry.