@cursor_ai: We're open-sourcing Mixture-of-Kittens (MoK), our MoE training megakernel for NVL72s. It fuses all Mixture-of-Experts c…
Summary
Cursor is open-sourcing Mixture-of-Kittens (MoK), a fused MoE training megakernel for NVIDIA NVL72 that runs up to 2.37x faster than the strongest public baselines.
View Cached Full Text
Cached at: 08/04/26, 06:13 PM
We’re open-sourcing Mixture-of-Kittens (MoK), our MoE training megakernel for NVL72s.
It fuses all Mixture-of-Experts communication and computation into a single, fully deterministic kernel, and runs up to 2.37x faster than the strongest public baselines. https://t.co/yHu5E6RXp9
Similar Articles
Mixture-of-Kittens: our open-source MoE megakernel for NVL72s (25 minute read)
Cursor is open-sourcing Mixture-of-Kittens (MoK), a production MoE training megakernel for NVL72s that fuses communication and computation, delivering a 1.41x end-to-end training throughput improvement for their Composer model.
40% speedup of MoE training with faster megakernel, by cursor, of all people (for B200s)
Cursor open-sources Mixture-of-Kittens (MoK), a deterministic MoE training megakernel for NVIDIA Blackwell GPUs that fuses computation and communication, delivering up to 2.37x speedup over baseline implementations.
@omarsar0: Banger compression paper from NVIDIA. (bookmark it) Bigger MoE models keep winning on quality, but serving them at inte…
NVIDIA's paper introduces Puzzle-75B-A9B, a compressed hybrid MoE model that doubles server throughput while preserving quality, enabling cost-effective deployment of large language models.
Nvidia released Nemotron 3.5 Super VL (5 minute read)
NVIDIA 发布 Nemotron 3.5 Super VL，一款 120B 参数的多模态混合专家（MoE）模型，每 token 激活 12B 参数，面向企业级智能体工作流，支持文本、图像、文档和视频理解，单 GPU 上吞吐量较同类开源模型提升高达 5 倍，并可通过 NeMo 进行后训练定制。
@samsja19: we made our training stack significantly faster by fusing computation and communication in the MoE layer, adding backwa…
Prime Intellect made DeepSeek's MegaMoE kernels trainable, fusing computation and communication in the MoE layer to achieve up to 55% faster MoE training than TorchTitan, with support for BF16, MXFP8, and FP8 on Blackwell and Hopper GPUs.