Moonshot AI Unveils Kim K3.1 Pro, Outperforming Claude Opus 5.5 and GPT-6 Astra
Moonshot AI today introduced Kim K3.1 Pro, a large language model (LLM) designed to surpass the performance of both Claude Opus 5.5 and GPT-6 Astra in several tasks.
The new LLM serves as a post-trained enhancement of the Kimi K3, which Moonshot AI released in July. This earlier model consists of 896 neural networks boasting a staggering 3 trillion parameters, with only 16 networks activated at a time to reduce inference costs.
To further optimize its hardware utilization, Moonshot AI incorporates Kimi Delta Attention within the Kim K3 framework. This technology, rooted in a memory-saving implementation called Gated DeltaNet, significantly decreases the RAM required by the model’s attention mechanism to analyze user prompts. It is expected that this feature has been retained in the Kim K3.1 Pro version.
According to the company, the post-training workflow of Kim K3.1 Pro demanded considerably more computational resources compared to previous projects. This latest iteration ran on an enhanced version of Moonshot AI’s post-training stack and included automated quality checks to improve overall infrastructure efficiency.
One notable area where Kim K3.1 Pro excels is in 3D modeling. The model differentiates itself by generating 3D objects through code snippets that define their geometry, rather than through conventional drawing methods. In a benchmarking exercise known as BenchCAD, Kim K3.1 Pro achieved an impressive score of 92.7%, positioning it just four percentage points behind Astra and Claude Opus 5.5.
Beyond static modeling, Kim K3.1 Pro is capable of animating 3D models. In an internal evaluation, the model was successfully used to create an animated simulation of human motion. Furthermore, Moonshot AI produced several 3D video games utilizing this model, which presents a considerable challenge due to the extensive coding involved, where even minor errors can result in crashes.
On the Terminal-Bench 4.0 coding benchmark, Kim K3.1 Pro scored 65.7%, outpacing Astra by more than 7%, yet trailing slightly behind Opus 5.5. Notably, it outperformed both competitors in the OfficeQA Pro v2 assessment, which evaluates LLM capabilities in executing complex financial analyses.
However, when it comes to cybersecurity benchmarks, Kim K3.1 Pro’s performance falls short compared to Astra and Opus 5.5. Moonshot AI indicated that its engineers intentionally excluded cybersecurity-related information from the model’s post-training dataset to mitigate safety risks, with plans to introduce more advanced vulnerability analysis features in future models.
The Kim K3.1 Pro is now accessible through Moonshot AI’s paid cloud services. The company has announced plans to open-source the model’s weights by November 4.