Cursor Research has open-sourced Mixture-of-Kittens (MoK), a specialized training kernel designed to optimize Mixture-of-Experts (MoE) models. This megakernel fuses communication and computation steps into a single deterministic process, reportedly achieving up to 2.37x higher throughput compared to existing baselines. MoK is designed for high-end hardware, specifically requiring NVIDIA Blackwell GB200 NVL72 or GB300 NVL72 racks, making it suitable for large-scale AI model development and cloud infrastructure providers. AI
IMPACT Enables faster training of large MoE models for organizations with high-end GPU infrastructure.
RANK_REASON Cursor Research open-sourced a specialized training kernel (MoK) for MoE models, which is a software tool rather than a new frontier model release.
- Cursor
- Mixture-of-Kittens
- Composer
- CUDA
- Cursor Research
- DeepSeek-V3
- GB300 NVL72
- Kimi 2.5
- NVIDIA Blackwell
- PyTorch
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →