PulseAugur
EN
LIVE 17:19:17

Cursor open-sources Mixture-of-Kittens training megakernel · 3 sources tracked

By PulseAugur Editorial · [3 sources] ·

Cursor has open-sourced Mixture-of-Kittens (MoK), an MoE training megakernel designed for NVL72s. This kernel fuses communication and computation for Mixture-of-Experts models, achieving up to 2.37x faster performance than existing baselines. MoK is currently used to train models across tens of thousands of GPUs at Cursor, improving their end-to-end training throughput by 1.41x. AI

IMPACT This open-source release aims to lower the barrier for AI research, enabling more labs to train models more efficiently.

RANK_REASON Cursor, an AI IDE, open-sourced a training megakernel.

Read on X — Cursor (AI IDE) →

AI-generated summary · Google Gemini · from 3 sources. How we write summaries →

Cursor open-sources Mixture-of-Kittens training megakernel · 3 sources tracked

COVERAGE [3]

  1. X — Cursor (AI IDE) TIER_1 English(EN) · cursor_ai ·

    Our hope is that this lowers the barrier to AI research, so more labs can train models efficiently.

    Our hope is that this lowers the barrier to AI research, so more labs can train models efficiently. Here's how we built it: https://t.co/EaC2z7Mw3a

  2. X — Cursor (AI IDE) TIER_1 English(EN) · cursor_ai ·

    MoK now powers training across tens of thousands of GPUs at Cursor.

    MoK now powers training across tens of thousands of GPUs at Cursor. In production, it raised end-to-end training throughput by 1.41x over our previous DeepEP-based stack.

  3. X — Cursor (AI IDE) TIER_1 English(EN) · cursor_ai ·

    We're open-sourcing Mixture-of-Kittens (MoK), our MoE training megakernel for NVL72s.

    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

RELATED ENTITIES

RELATED TOPICS