# MiniMax 研究员谈稀疏注意力与原生多模态训练

- 来源：MiniMax (official) (@MiniMax_AI)
- 发布时间：2026-07-01 14:44
- AIHOT 分数：45
- AIHOT 链接：https://aihot.virxact.com/items/cmr1q263c00z0sl8zhops7vtb
- 原文链接：https://x.com/MiniMax_AI/status/2072209671088443706

## AI 摘要

MiniMax 强化学习研究负责人姜松岩在 AI Engineer 大会第二天参加炉边谈话，与 Thom Wolf 探讨了稀疏注意力、从训练第一天就融入的原生多模态方案，以及开放权重对 AI 发展的意义。她强调分享了团队在这些方向的工作，并重申相信开放权重的重要性。

## 正文

Day two of @aiDotEngineer started with a conversation anyone serious about open weights should be paying attention to.

@olive_jy_song， research lead RL， joined @Thom_Wolf to dig into sparse attention， native multimodal training from day zero， and why open-weights matter for where AI is headed.

Grateful to @swyx and the AI Engineer team for creating spaces for technical exchanges at this level.

### 引用推文

> Olive Song：Backstage and onstage with @Thom_Wolf and @swyx . I really enjoyed the fireside chat! Thanks for having me back at @aiDotEngineer! And always proud to be part o...