Moonshot AI Unveils Kimi K3.1 Pro, a 3-Trillion-Parameter Model Rivaling OpenAI and Anthropic
The Chinese AI firm has released an upgraded version of its Kimi K3 model that matches or exceeds competitors in several specialized tasks, particularly 3D modeling and financial analysis.
Moonshot AI has unveiled Kimi K3.1 Pro, a large language model that can perform some tasks better than Claude Opus 5.5 and GPT-6 Astra. The new model represents a post-trained refinement of Kimi K3, which the Chinese artificial intelligence company released in July.
Kimi K3 is built on 896 individual neural networks containing 3 trillion parameters in total. The architecture activates only 16 of those neural networks at any given time, a design choice intended to keep inference costs manageable.
Beyond its sparse activation approach, Moonshot AI has implemented additional hardware optimizations. Kimi K3 incorporates Kimi Delta Attention, a proprietary version of Gated DeltaNet that reduces the RAM requirements for the model's attention mechanism when processing user prompts. This feature has presumably carried forward to Kimi K3.1 Pro.
The post-training process for the new model demanded substantially more computational resources than Moonshot AI's earlier efforts. The company deployed an enhanced iteration of its post-training infrastructure, which now includes automated quality assurance systems and improved underlying efficiency.
Performance Across Specialized Domains
Kimi K3.1 Pro demonstrates particularly strong capabilities in 3D modeling. The model can generate 3D objects by writing code that defines their geometry numerically rather than requiring manual design. On BenchCAD, an industry benchmark for 3D model coding, Kimi K3.1 Pro achieved 92.7%, placing it within four percentage points of both Astra and Claude Opus 5.5.
Beyond static geometry, Kimi K3.1 Pro can animate 3D models. Moonshot AI demonstrated this capability by using the model to generate animated simulations of human motion and to create multiple 3D video games, a particularly challenging task because games contain extensive code where even minor bugs can cause failures.
On Terminal-Bench 4.0, a standard coding benchmark, Kimi K3.1 Pro scored 65.7%, outpacing Astra by more than 7 percentage points but trailing slightly behind Opus 5.5. The model also exceeded both competitors on OfficeQA Pro v2, a test measuring LLMs' performance on complex financial analysis tasks.
Performance on cybersecurity benchmarks represents a notable weakness. Kimi K3.1 Pro's scores on popular security tests significantly lag behind those of Astra and Opus 5.5. Moonshot AI deliberately excluded cybersecurity-related information from the model's post-training dataset to mitigate safety risks, though the company intends to add more sophisticated vulnerability analysis capabilities to future versions.
Availability and Open-Source Plans
Kimi K3.1 Pro is accessible immediately through Moonshot AI's commercial cloud services. The company has committed to open-sourcing the model's weights by November 4.
Read more
Identity Security Must Adapt as AI Agents Operate Across Enterprise Systems
Okta's Oktane event revealed that securing AI agents requires shared architecture, cross-vendor cooperation and runtime controls that extend identity management beyond initial login.
Managed Services Become Critical to Neocloud Success as GPU Providers Scale
Smaller AI infrastructure providers are increasingly relying on hardware partner support to keep their GPU clusters operational and profitable after launch, with managed services becoming a key differentiator in the competitive neocloud market.
Why AI ROI Remains Elusive: Breaking Out of the Productivity Trap
Companies are discovering that deploying AI tools in isolation—without overhauling the business processes around them—delivers disappointing returns. The gap between hype and reality is what some call AI's "valley of despair."