Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks
Cursor Research has open-sourced Mixture-of-Kittens (MoK), a deterministic mixture-of-experts training kernel that accelerates model training by up to 2.37x on NVIDIA GB300 NVL72 racks.
- MoK is a deterministic mixture-of-experts training kernel that fuses communication and computation into a single operation.
- It delivers up to 2.37x speed improvements on NVIDIA GB300 NVL72 racks compared to public baselines.
- Requires Blackwell SM100 or SM103 GPUs, limiting accessibility to organizations with NVL72 capacity.
- Cursor Research has open-sourced MoK to encourage broader adoption and innovation in MoE training.
Cursor Research has open-sourced Mixture-of-Kittens (MoK), a deterministic mixture-of-experts (MoE) training megakernel designed to optimize performance on NVIDIA GB300 NVL72 racks. The kernel consolidates all MoE communication and computation into a single, deterministic operation, significantly reducing overhead and improving training speed.
According to Cursor, MoK achieves up to 2.37x faster training compared to the strongest public baseline when running on GB300 NVL72 systems. This performance boost is tied to the use of Blackwell SM100 or SM103 GPUs, which are required for MoK to function. The open-source release targets developers working with large-scale MoE models, particularly those leveraging NVIDIA's latest hardware.
The move underscores Cursor's commitment to advancing efficient training methodologies for large language models. By releasing MoK under an open-source license, the company aims to foster collaboration and innovation within the AI research community.
Provides a high-performance, deterministic MoE training kernel for large-scale models.
Enables faster model training on NVIDIA GB300 systems, reducing computational costs.
Highlights advancements in AI training efficiency, potentially increasing value in AI infrastructure.
Demonstrates progress in optimizing AI training for next-generation hardware.
- Mixture-of-Experts (MoE)
- A machine learning model architecture that uses multiple specialized sub-models (experts) and selectively activates only a subset for each input, improving efficiency.
- Megakernel
- A single, highly optimized kernel that combines multiple operations to reduce overhead and improve performance.
- GB300 NVL72
- NVIDIA's high-performance computing system designed for large-scale AI workloads, featuring Blackwell architecture GPUs.
AI ToolsYour agent writes Python. The Ruby rule cuts that by a third.
AI ToolsThe Channel Gap: Why Your LLM Judge is Blind in One Eye
AI Toolsclaude -p: what headless Claude Code actually loads (and when --bare is the right call)
AI ToolsBaseten on Hugging Face Inference Providers 🔥
AI ToolsResize One Image into 6 Social Media Formats Automatically Using Cloudinary Claimable Clouds
AI ResearchWeatherNext: AI model achieves breakthrough in forecasting cyclones
DeepMind introduced WeatherNext, an AI system that markedly improves cyclone track and intensity predictions, extending forecast lead times by several days.
US Senate Commerce approves KOSA, children's AI safety bills - IAPP
The US Senate Commerce Committee has approved two bills focused on AI safety for children. The bills aim to regulate AI systems and protect children's data.
Artificial intelligence enters Italy’s national security agenda - Decode39
Italy has added artificial intelligence to its national security agenda, marking a significant development in the country's approach to AI. This move is expected to have implications for the nation's defense and security strategies.
Powering the ballot: Why AI’s energy footprint is the ultimate midterm election issue - Route Fifty
AI’s growing energy demands are becoming a key issue in the US midterm elections, raising questions about sustainability and infrastructure.
Accelerating Biomedical Innovation with AI through Collaborative Iteration - Wyss Institute at Harvard
The Wyss Institute at Harvard is leveraging AI to accelerate biomedical innovation through collaborative iteration. Researchers are using AI to analyze and improve medical devices and treatments.
DeepSeek invests $20.8 million in Unitree's Shanghai IPO - Reuters
DeepSeek has committed $20.8 million to Unitree's upcoming Shanghai IPO, signaling strong investor confidence in the robotics firm.