China's Moonshot AI Launches Kimi K3.1 Pro, a 3-Trillion-Parameter Model That Beats GPT-6 Astra on Coding
Moonshot AI released Kimi K3.1 Pro on 23 October 2026, a post-trained upgrade to its Kimi K3 model that outperforms GPT-6 Astra on key coding and financial benchmarks and will be open-sourced by 4 November 2026.
Chinese AI company Moonshot AI unveiled Kimi K3.1 Pro on 23 October 2026, a post-trained upgrade of its Kimi K3 model, which was first released in July. The underlying Kimi K3 architecture contains 896 individual neural networks totalling 3 trillion parameters, but activates only 16 of them at any given time to keep inference costs low. A custom memory-saving feature called Kimi Delta Attention, built on a technology known as Gated DeltaNet, reduces the RAM the model's attention mechanism needs when interpreting user prompts. Moonshot says the post-training process consumed significantly more compute than previous projects and ran on an upgraded pipeline with automated quality checks and improved infrastructure efficiency.
On the Terminal-Bench 4.0 coding benchmark, K3.1 Pro scored 65.7%, more than 7 percentage points ahead of GPT-6 Astra and slightly behind Claude Opus 5.5. On the BenchCAD 3D-modelling benchmark it reached 92.7%, within four percentage points of both rivals. Beyond code generation, the model can animate 3D objects; Moonshot used it internally to simulate human motion and generate multiple 3D video games. K3.1 Pro also outperformed Astra and Opus 5.5 on OfficeQA Pro v2, a complex financial analysis test. Cybersecurity benchmarks are a noted weakness: Moonshot deliberately removed cybersecurity data from the post-training dataset for safety reasons and plans to add stronger vulnerability analysis to future models.
Kimi K3.1 Pro is available now through Moonshot AI's paid cloud services, with open-source model weights planned for release by 4 November 2026.
