--- schema: 1 type: video title: >- Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song description: >- In this conversation, Olive Song, who leads reinforcement learning at MiniMax, opens up the stack behind the company's open weight models and the infrastructure that serves them. Her starting point is a belief in open source: put the weights out, let builders optimize on them, and share the capability widely. From there she walks through what it takes for a model to land well, from agentic coding that also understands and builds games, to computer use problems trained with RL against environments like OS World. Much of the discussion is the unglamorous engineering that makes a launch real. When a model ships, the team wants the inference stack ready on day zero, which means writing and tuning GPU kernels, working through benchmarks like a parallel kernel bench, and threading optimization into everything from KV cache handling to routing. Song also talks through multimodality and the training pitfalls that come with it, such as text and vision collapsing after training unless you train both modalities together, and reflects on longer horizon tasks like replicating a twelve hour run. She closes optimistic that open models are closing the gap faster than a year ago would have suggested. Speaker info: - https://x.com/olive_jy_song Timestamps: 0:00 - Introducing the RL lead at MiniMax 1:17 - Why open source and open weights 3:45 - What builders are doing with the model 4:11 - A model that builds games 5:12 - Computer use and OS World 5:49 - Writing GPU kernels 6:25 - Parallel kernel bench 7:28 - A day zero inference stack 9:34 - Optimization across the stack 10:47 - Multimodality and training collapse 14:10 - Replicating a twelve hour run 17:22 - Where open models go next canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/agents-at-scale-inside-minimax-s-model-and-the-infr-624dd1b2bc33 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/agents-at-scale-inside-minimax-s-model-and-the-infr-624dd1b2bc33.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T06:29:35.742Z' language: null children: [] videoId: AVMr9PMINyo durationSeconds: 1214 publishedAt: '2026-07-31T03:00:14.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 77 endSeconds: 88 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=77' heldReason: rights - ordinal: 1 text: null startSeconds: 88 endSeconds: 234 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=88' heldReason: rights - ordinal: 2 text: null startSeconds: 257 endSeconds: 267 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=257' heldReason: rights - ordinal: 3 text: null startSeconds: 267 endSeconds: 312 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=267' heldReason: rights - ordinal: 4 text: null startSeconds: 324 endSeconds: 358 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=324' heldReason: rights - ordinal: 5 text: null startSeconds: 369 endSeconds: 393 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=369' heldReason: rights - ordinal: 6 text: null startSeconds: 405 endSeconds: 451 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=405' heldReason: rights - ordinal: 7 text: null startSeconds: 451 endSeconds: 462 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=451' heldReason: rights - ordinal: 8 text: null startSeconds: 462 endSeconds: 582 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=462' heldReason: rights - ordinal: 9 text: null startSeconds: 582 endSeconds: 594 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=582' heldReason: rights - ordinal: 10 text: null startSeconds: 647 endSeconds: 659 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=647' heldReason: rights - ordinal: 11 text: null startSeconds: 659 endSeconds: 852 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=659' heldReason: rights - ordinal: 12 text: null startSeconds: 659 endSeconds: 852 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=659' heldReason: rights - ordinal: 13 text: null startSeconds: 863 endSeconds: 1049 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=863' heldReason: rights - ordinal: 14 text: null startSeconds: 863 endSeconds: 1049 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=863' heldReason: rights - ordinal: 15 text: null startSeconds: 863 endSeconds: 1049 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=863' heldReason: rights - ordinal: 16 text: null startSeconds: 1049 endSeconds: 1061 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=1049' heldReason: rights - ordinal: 17 text: null startSeconds: 1061 endSeconds: 1214 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=1061' heldReason: rights - ordinal: 18 text: null startSeconds: 1061 endSeconds: 1214 watch: 'https://liquidgold.cc/watch/AVMr9PMINyo?t=1061' heldReason: rights --- # Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song In this conversation, Olive Song, who leads reinforcement learning at MiniMax, opens up the stack behind the company's open weight models and the infrastructure that serves them. Her starting point is a belief in open source: put the weights out, let builders optimize on them, and share the capability widely. From there she walks through what it takes for a model to land well, from agentic coding that also understands and builds games, to computer use problems trained with RL against environments like OS World. Much of the discussion is the unglamorous engineering that makes a launch real. When a model ships, the team wants the inference stack ready on day zero, which means writing and tuning GPU kernels, working through benchmarks like a parallel kernel bench, and threading optimization into everything from KV cache handling to routing. Song also talks through multimodality and the training pitfalls that come with it, such as text and vision collapsing after training unless you train both modalities together, and reflects on longer horizon tasks like replicating a twelve hour run. She closes optimistic that open models are closing the gap faster than a year ago would have suggested. Speaker info: - https://x.com/olive_jy_song Timestamps: 0:00 - Introducing the RL lead at MiniMax 1:17 - Why open source and open weights 3:45 - What builders are doing with the model 4:11 - A model that builds games 5:12 - Computer use and OS World 5:49 - Writing GPU kernels 6:25 - Parallel kernel bench 7:28 - A day zero inference stack 9:34 - Optimization across the stack 10:47 - Multimodality and training collapse 14:10 - Replicating a twelve hour run 17:22 - Where open models go next Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/agents-at-scale-inside-minimax-s-model-and-the-infr-624dd1b2bc33 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.