Videos AVMr9PMINyo

Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song

index_state ready data_status ok

AI Engineer· published 2026-07-31· 0:20:14· en-US· indexed 2026-08-10 19:37

Open on YouTube

51 shot(s).

keyframes kept every frame deduplicated

What was stored

cues: 250; whisperx 250
chunks: 36; from 250 cues
keyframes: 48; kept of 51 captured
frames with text: 48; 118 lines read
chapters: 12; from the source metadata
keyframe bytes: 5.3 MB; word timings on 250 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage	state	model	started	took
`fetch`	done	—	2026-08-09 21:47	0s
`stt`	done	—	2026-08-09 02:52	24s
`chunk`	done	—	2026-08-09 02:53	0s
`text_embed`	done	—	2026-08-10 19:37	1s
`keyframe`	done	—	2026-08-09 02:53	2m 53s
`ocr`	done	—	2026-08-09 02:55	7s
`frame_embed`	done	—	2026-08-10 19:37	8s

Frames, and what the machine read

0:02 #0 done2 line(s)

shot 0·sharpness 459.9
1. AlEngineer0.96
2. World's Fair1.00
0:03 #1 done2 line(s)

shot 1·sharpness 662.1
1. AlEngineer0.95
2. World's Fair0.99
0:10 #2 done24 line(s)

shot 2·sharpness 2744.8
1. LAB & PLATINUM SPONSORS0.99
2. Amazon AGI Lab0.98
3. ANTHROP\C1.00
4. Google DeepMind1.00
5. MINIMAX0.94
6. OpenAI0.92
7. Akamai1.00
8. arize1.00
9. aws1.00
10. Braintrust bright data0.98
11. B1.00
12. Browserbase1.00
13. docker1.00
14. :neo4j0.92
15. ORACLE1.00
16. PayPal1.00
17. qodo1.00
18. reducto1.00
19. Sonar1.00
20. Makers of1.00
21. togetherai1.00
22. Unblocked1.00
23. WorkOS1.00
24. SonarQube1.00
0:17 #3 done2 line(s)

shot 3·sharpness 360.6
1. AlEngineer1.00
2. World's Fair0.98
0:40 #4 done2 line(s)

shot 4·sharpness 364.5
1. AlEngineer1.00
2. World's Fair0.98
1:19 #5 done2 line(s)

shot 5·sharpness 365.0
1. AlEngineer1.00
2. World's Fair0.99
1:30 #6 done2 line(s)

shot 6·sharpness 346.1
1. AlEngineer1.00
2. World's Fair0.98
2:05 #7 done2 line(s)

shot 7·sharpness 344.3
1. AlEngineer1.00
2. World's Fair0.98
2:20 #8 done2 line(s)

shot 8·sharpness 358.5
1. AlEngineer1.00
2. World's Fair0.98
2:55 #9 done2 line(s)

shot 9·sharpness 228.7
1. AlEngineer0.96
2. World's Fair1.00
3:27 #10 done2 line(s)

shot 10·sharpness 221.7
1. AlEngineer0.97
2. World's Fair1.00
3:55 #11 done2 line(s)

shot 11·sharpness 346.0
1. AlEngineer1.00
2. World's Fair0.99
4:06 #12 done2 line(s)

shot 12·sharpness 348.6
1. AlEngineer0.99
2. World's Fair0.97
4:31 #13 done2 line(s)

shot 13·sharpness 238.4
1. AlEngineer0.96
2. World's Fair0.97
4:51 #14 done2 line(s)

shot 14·sharpness 231.2
1. AlEngineer0.96
2. World's Fair1.00
5:26 #15 done2 line(s)

shot 15·sharpness 227.9
1. AlEngineer0.97
2. World's Fair1.00
5:41 #16 done2 line(s)

shot 16·sharpness 255.7
1. AlEngineer0.99
2. World's Fair1.00
6:18 #17 done2 line(s)

shot 17·sharpness 236.3
1. AlEngineer0.99
2. World's Fair1.00
6:41 #18 done2 line(s)

shot 18·sharpness 238.5
1. AlEngineer0.98
2. World's Fair1.00
6:59 #19 done2 line(s)

shot 19·sharpness 236.3
1. AlEngineer0.98
2. World's Fair0.97
7:38 #20 done2 line(s)

shot 20·sharpness 231.0
1. AlEngineer0.98
2. World's Fair0.99
7:52 #21 done2 line(s)

shot 21·sharpness 229.2
1. AlEngineer0.97
2. World's Fair1.00
8:26 #22 done2 line(s)

shot 22·sharpness 243.1
1. AlEngineer0.97
2. World's Fair1.00
8:37 #23 done2 line(s)

shot 23·sharpness 236.0
1. AlEngineer0.99
2. World's Fair1.00

Transcript

250 cues· 3,628 words· 19,400 chars

chunk 0 · 0:12–0:56 · 157 words · 802 chars
0:12 This is a discussion that I'm particularly excited about because the field is moving so fast and we have two people that kind of have this unique vantage point on the field.
0:22 And what I want to do is kind of ask questions to see if we can learn from that.
0:26 So I want to start off with intros, talk a little bit about your role, what you're thinking about, what you're working on.
0:31 Maybe Dan, if you can go first.
0:33 Hey, everyone.
0:34 I'm Dan.
0:34 I'm the VP of Kernels at Together AI.
0:37 I lead inference, GPU optimization, trying to figure out how to use GPUs most effectively to serve AI models.
chunk 1 · 0:45–1:29 · 146 words · 759 chars
0:45 Yeah, so one of the things that I wanted to dive in with Dan about is new model drops, what is everything that goes on behind the scenes to serve it so that everybody here, there's a lot of builders here that can use it.
0:57 Olive, I want to throw it over to you to talk about your role and what you're focusing on.
1:01 Yeah, I'm Olive, and I am the research lead of RL at Minimax.
1:06 And I am responsible for the final training of the model and the shipping of the model, so basically everything before the infrastructure.
1:13 Awesome.
1:14 So maybe I want to start off, this panel is focusing on open source.
chunk 2 · 1:18–2:01 · 121 words · 680 chars
1:18 I wanted to start off with, this is your strongest model yet, Minimax M3.
1:23 Why open source it?
1:24 What's the idea behind that as a company as you're releasing these models?
1:29 We do believe that the open source community as a whole is very strong and powerful.
1:35 While we open source the model, everyone can use it.
1:38 So it aligns with our mission that we want to have intelligence with everyone.
1:42 And also different developers can contribute to the model through feedback, through their own PRs, and we can build the models even stronger.
chunk 3 · 1:50–2:32 · 125 words · 693 chars
1:50 And also, for example,
1:53 then you will be able to optimize on our open-weight model and make it inference faster and then serve better for everyone.
2:00 Yeah.
2:01 Yeah, we're big believers in open-source set together.
2:03 And yeah, I think we've been following you guys for a while, I think, from way older Minimax models.
2:11 So seeing M3 and seeing how far it's come is really impressive and really great.
2:16 So I wanted to kind of pick on this a little bit more.
2:20 Can you explain?
chunk 4 · 2:21–3:04 · 135 words · 744 chars
2:21 So we've got the model creators themselves, Minimax.
2:23 We've got experts on the inference side of things.
2:26 How did this partnership come to be?
2:28 So they launch an open source model and we're now distributing it.
2:31 I checked this morning.
2:32 We have the lion's share of token usage for Minimax M3.
2:38 How does this partnership come to be and how do we serve a model like this at scale?
2:41 Yeah, yeah, great question.
2:42 So at Together, I think one of the things that we're really interested in is how do you make intelligence abundant?
2:49 So how do you get more tokens for more people to do more useful things and get all these capabilities into more people's hands?
chunk 5 · 2:58–3:39 · 124 words · 654 chars
2:58 So we follow all the open models very closely.
3:02 I don't remember when exactly we started partnering.
3:04 Oh, actually, I think I do know this.
3:06 We had a car event in Las Vegas sometime last year, and someone from Minimax came, and there he was like, guys, you really gotta serve our next model.
3:14 It's gonna be really, really great.
3:16 So I think from there, we started talking.
3:18 We were serving Minimax 2.5.
3:21 and I think 2.7 for a while.
3:24 And then when leading up to the launch of M3, we were quite excited about it, I think.
chunk 6 · 3:30–4:13 · 150 words · 801 chars
3:30 We were seeing the usage and what people were doing with it.
3:34 It was really quite exciting.

Chapters

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