Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax

Read the talk

Why Agents Need More Context: MiniMax M3’s Attention, Multimodal Training and Research Workflows

Thomas Wolf and Olive Song discuss how long context, selective attention and training with text and vision fit together—and how model feedback and agent harnesses feed into the next round of research.

From a talk by Thomas Wolf and Olive Song

At a glance

Ideas worth remembering

  • Long context serves an agent’s accumulating interactions and tool responses. MiniMax Sparse Attention makes that context more tractable by selecting blocks with an index branch and computing attention over the selection.

  • Native multimodal training addresses problems MiniMax observed when adding vision later, but introduces its own stability challenge. Song credits vision-component work, interleaved data, cleaning, masking and reward modeling with preventing collapse.

  • Multimodal understanding becomes especially useful when it feeds subsequent action: presentations, reports and long videos can supply information for tool-using agents. The interview proposes these workflows without demonstrating their reliability.

  • MiniMax describes several routes from observed behavior to better models: internal evaluations and research projects, community feedback and pull requests, and automated research harnesses that help accelerate iteration.

  • The parameter figures are not fully consistent: the opening gives approximately 400 billion total and 20 billion activated, while the later discussion gives 428 billion total and 23 billion active. The supported distinction is between total and active parameters; the exact counts remain uncertain.

  • Song’s closing emphasis is on multi-agent systems and model routing, both to handle more complex tasks and to expose the boundaries of individual models’ capabilities.

An open-model conversation

Selected presentation frame from Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax at 51 seconds
An open-model conversation

After the musical introduction, Hugging Face co-founder and chief science officer Thomas Wolf welcomes Olive Song to discuss MiniMax. He frames the conversation around competition among leading open-source model teams and introduces M3 as MiniMax’s recent release. His leaderboard comparisons establish the interview’s context, rather than providing a detailed evaluation of the models.

0:120:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:01 · section reference included

Coding, vision and a million tokens

Selected presentation frame from Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax at 194 seconds
Coding, vision and a million tokens

Song introduces M3 with approximate figures of 400 billion total parameters and 20 billion activated parameters. These are distinct quantities: the model’s total parameter count is larger than the count activated during its computation. She emphasizes the combination of coding performance, image and video understanding, and a one-million-token context window, enabled by an architecture called MSA, or MiniMax Sparse Attention.

The design goal is to bring these capabilities together for future applications. Coding and agentic behavior address what the model can do; multimodal understanding broadens what it can receive; longer context expands how much information it can work with at once. Wolf stresses that the million-token window is intended to be functional, then asks how the attention architecture makes that length efficient. The exchange supplies a capability claim, but no benchmark results or measurement procedure for that claim.

2:342:36
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:34 · section reference included

From reading a book to retaining an agent’s history

Selected presentation frame from Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax at 338 seconds
From reading a book to retaining an agent’s history

Song traces MiniMax’s long-context work to earlier models that she says could perform tasks with ten million tokens of context. Her example is supplying a book and asking for a review. She distinguishes that use from an agentic model: accepting a large input does not by itself establish that a model can manage repeated actions and interactions.

An agent’s context grows through interaction. It receives user messages, acts in an environment, gets tool responses and proceeds through multiple rounds. Song argues that a short window cannot hold enough of this accumulating information for complex tasks. Restoring longer context therefore serves a concrete purpose: keeping more of the material generated during the task available while the agent continues working.

MiniMax Sparse Attention separates selection from attention computation. An index branch identifies which parts of the context matter more; a sparse-attention branch then computes over the selected blocks. The efficiency mechanism is selective computation over context blocks. That makes the quality of selection consequential, although Song does not specify the selection rule, block sizes or how selection errors affect performance. She describes the simple architecture as a foundation for scaling both context length and model size.

4:124:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:12 · section reference included

Efficiency remains an architecture problem

Selected presentation frame from Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax at 463 seconds
Efficiency remains an architecture problem

Wolf places sparse attention in the history of efforts to address attention’s quadratic, or N-squared, computation. He recalls work on linear attention and the subsequent prominence of FlashAttention, then welcomes renewed investigation into attention’s underlying design. His comparison with GPT-2’s 1,024-token window illustrates how expectations for context length have expanded.

When Wolf raises trillion-token attention, Song treats it as a research direction. She says extreme context lengths would require work on architecture together with hardware. She also sees room for further architecture and inference optimization, particularly where applications need strong capabilities and efficient operation. No concrete trillion-token implementation or quantified efficiency gain is offered.

The architecture’s origin turns the discussion toward research access. Song credits an intern with designing it and contrasts MiniMax’s openness to contributions with labs where interns may lack access to data or substantive work. The example connects technical invention to who can use the resources needed to investigate an idea.

5:576:01
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:57 · section reference included

Turning observed weaknesses into research projects

Selected presentation frame from Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax at 555 seconds
Turning observed weaknesses into research projects

Song describes an internal process built on shared infrastructure. Researchers can use the model, devise their own evaluations, identify weaknesses and propose improvements. Interested colleagues join a project, work on it for weeks or months, and incorporate successful results into final training and a subsequent release. Model access and evaluation make a proposed improvement concrete before it becomes part of training.

Architecture projects can require a longer investigation. Song lists research, experiments and even redoing pretraining evaluations as necessary work. This gives the process room for deeper changes whose consequences cannot be established through a quick trial.

8:568:59
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:24 · section reference included

Training text and vision from the first step

Selected presentation frame from Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax at 747 seconds
Training text and vision from the first step

Song calls MiniMax’s approach native multimodality. She contrasts it with completing text pretraining, adding adapters and then training vision understanding. In MiniMax’s experiments, she says, the later addition harmed text performance while vision performance did not converge well. Her explanation is that the model had already converged toward text understanding, making the subsequent adjustment less effective.

Introducing vision partway through pretraining presents another difficulty: sensitivity to the training recipe. Song names architecture, data mixtures and learning rates as variables that change the result. A recipe that works in one experiment may therefore fail to carry its conclusions to a larger model. Her concern is both control of the experiment and scalability of what researchers learn from it.

Training multimodally from the first step avoids that later transition, but Song acknowledges a serious failure mode: text and vision understanding can collapse after only a few training steps. She says MiniMax addressed this through work on the vision component and the training data. The data remain interleaved, retaining images and videos in their natural context, alongside cleaning and masking. She also credits reward modeling. These are the ingredients she identifies; the interview does not provide the detailed recipe or isolate each ingredient’s contribution to stability.

10:3710:40
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:00 · section reference included

Larger models and interfaces people can use

Wolf asks about scaling beyond a trillion parameters, giving figures of 428 billion total parameters and 23 billion active parameters for the model under discussion. Song says MiniMax intends to pursue larger models because some tasks perform poorly with smaller parameter counts. This is an ambition and a motivation, without a specified future architecture or evidence showing which tasks require that scale.

The conversation then moves from model scale to MiniMax’s applications. Wolf asks whether apps and their data came first. Song describes the company as model-led from its beginning: the founding ambition was a model that could understand and output across modalities. That broad ambition should be distinguished from the specific capabilities demonstrated or described for any individual release.

Apps followed as a way to make model capabilities accessible. Song argues that people cannot all be expected to use an API, so useful interfaces, interactions and application scenarios matter. She reports reach of more than 300 million people across around 200 countries, and over a million companies. The figures describe the scale she attributes to MiniMax’s products; the interview does not define activity levels or the measurement period.

12:5512:57
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:53 · section reference included

Open releases as a source of improvements

Wolf raises the tension between open releases and the need for revenue. Song responds with the research team’s intention to keep open-sourcing models and the value it receives from the community. Performance feedback and pull requests inform later versions. Her answer explains a research benefit of openness, while leaving the revenue model and the relationship between app-specific models and public releases unresolved.

Asked what feedback would help most, Song welcomes reports of failures, especially in multimodal behavior. She acknowledges flaws in the current combination of capabilities and describes improvement as ongoing. Feature requests also matter: she gives thinking effort as an example of something users have requested. This is an invitation to influence future models, rather than a commitment to a particular implementation or delivery date.

15:2215:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:10 · section reference included

Understanding presentations and videos before acting

Wolf and Song agree that multimodality in coding agents remains underexplored. Song’s examples show the broader opportunity: reading a PowerPoint presentation, interpreting a report that is not well structured, or understanding a long video and then using tools. The proposed workflow connects perception to action, with the content supplying information the agent needs for its subsequent task.

Wolf makes the idea concrete by asking whether an agent could watch his YouTube tutorial and learn how to use the coding tools he describes. Song says she thinks so. It is a plausible application in the discussion, but the exchange does not establish that a particular tutorial was tested or that the resulting tool use was successful.

17:2917:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:29 · section reference included

Research harnesses and the move toward multiple agents

Song says MiniMax builds its own research harnesses and uses them to automate many workflows. She mentions kernel optimization and building data as examples of work models can support. The harness combines model capabilities with a workflow so they can assist daily research and accelerate iteration. She reports that M3 is already helping develop its next version, though the exchange does not establish how much of that development is automated or quantify the speedup.

For the closing question, Song identifies multi-agent systems and model routing as the direction she finds most exciting. She sees them as ways to support more complex tasks and reveal what models can and cannot do. This puts application design alongside model development: combining and routing models can expand what a system accomplishes while exposing capability limits. She offers a direction rather than a specific coordination architecture, and the interview ends with thanks and applause.

18:2718:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:27 · section reference included

Read the complete timestamped transcript
  1. 0:01

    [music]

  2. 0:12

    >> Joining us on stage is the co-founder

  3. 0:14

    and chief science officer at Hugging

  4. 0:16

    Face, Thomas Wolf.

  5. 0:20

    >> [music]

  6. 0:26

    [music]

  7. 0:32

    >> Hello everyone.

  8. 0:33

    Hello Olive, nice to have you on stage.

  9. 0:36

    >> Hi, nice to meet you. Thanks for having

  10. 0:37

    me, yeah.

  11. 0:38

    >> So I think you're on for a treat today

  12. 0:40

    because you just saw a GLM

  13. 0:43

    uh which is current number

  14. 0:46

    two on the artificial intelligence

  15. 0:48

    leaderboard. I take Fable out because

  16. 0:50

    nobody can use it. And now we have

  17. 0:52

    number four. So you basically you will

  18. 0:53

    have all the top models, at least the

  19. 0:55

    top open source model in a row. And

  20. 0:57

    we're very lucky to have

  21. 0:59

    Olive who has a

  22. 1:01

    pretty amazing path in life.

  23. 1:04

    Uh so she came to the US, Pennsylvania.

  24. 1:06

    She was studying, doing PhD at uh

  25. 1:09

    NYU uh in the lab of Jan LeCun

  26. 1:13

    working on J Pa, but we decided we won't

  27. 1:15

    talk about J Pa today, right?

  28. 1:17

    Something for another day. Um and then

  29. 1:20

    instead of joining Hugging Face, which

  30. 1:21

    was in New York also at that time she

  31. 1:24

    decided to go join MiniMax.

  32. 1:28

    So for those who who maybe don't know

  33. 1:30

    all the all the neo labs around the

  34. 1:33

    world and you're you're forgiven because

  35. 1:35

    I think there's like 64 neo labs right

  36. 1:37

    now. MiniMax is one of the top of what

  37. 1:40

    we call the AI dragons in China.

  38. 1:43

    So these are the new there's there's

  39. 1:45

    Deep Seek which is very well known now,

  40. 1:47

    Moonshot who does Kimi Z and GLM that

  41. 1:50

    you just saw and now we have a MiniMax.

  42. 1:53

    They're all extremely good, extremely a

  43. 1:55

    team

  44. 1:56

    uh fighting for the first spot.

  45. 1:59

    Uh so the the the latest uh release of

  46. 2:01

    MiniMax was M3 uh just earlier earlier

  47. 2:05

    in June, which was the the top model at

  48. 2:07

    the time, top open-source model.

  49. 2:09

    Uh very impressive. There's a lot of

  50. 2:11

    very interesting things about this

  51. 2:12

    model, so we'll quickly dive in them.

  52. 2:15

    And then talk a little bit about uh

  53. 2:17

    what's what's what's specific about

  54. 2:18

    MiniMax, what's what's great there.

  55. 2:22

    So uh maybe Olive to to start a little

  56. 2:24

    bit. Can you Can you give us, you know,

  57. 2:26

    a

  58. 2:27

    a little bit of your your view of of M3,

  59. 2:30

    what you like about this model, how was

  60. 2:32

    the release?

  61. 2:33

    >> Mhm.

  62. 2:34

    Yeah, M3 we released M3 earlier this

  63. 2:36

    month, and it is a smaller model with

  64. 2:40

    400 around 400 billion total parameters

  65. 2:43

    and 20 billion activated. Um

  66. 2:46

    but it is very capable in terms of both

  67. 2:48

    coding performances, and also it

  68. 2:51

    understands vision. So um

  69. 2:54

    that's uh what

  70. 2:56

    open-source models don't usually have.

  71. 2:58

    It's that they can

  72. 3:00

    the model can only deal with coding, but

  73. 3:03

    it can also understand videos, um

  74. 3:05

    images, and it has a super uh long

  75. 3:07

    context of 1 million.

  76. 3:09

    Um with our new architecture called MSA,

  77. 3:12

    MiniMax Sparse Attention. So we

  78. 3:15

    we really put these three things

  79. 3:17

    together uh because we know that they

  80. 3:21

    are they will be very important in

  81. 3:23

    future AI applications. Coding

  82. 3:25

    capabilities, agentic capabilities,

  83. 3:28

    longer context, and multimodal

  84. 3:30

    understanding. Um yeah, I think that

  85. 3:33

    would be very interesting about the

  86. 3:34

    model.

  87. 3:35

    >> Yeah, so so there's a lot to unpack

  88. 3:37

    unpack in this model, and it's um

  89. 3:39

    it's it's still, I think, the only top

  90. 3:41

    five model open-source model that is

  91. 3:43

    actually multimodal, so we need to talk

  92. 3:45

    about that. But maybe first about the

  93. 3:47

    long context, because that was also the

  94. 3:48

    first one that really had this real 1

  95. 3:51

    million token long context is actually

  96. 3:53

    functional and you guys had also the the

  97. 3:56

    Minimax pass attention

  98. 3:59

    which is this one technique to to make

  99. 4:00

    that efficient that you also published

  100. 4:03

    and and share extensively. So can you

  101. 4:05

    can you talk a little bit about this?

  102. 4:07

    Maybe how the project went from

  103. 4:09

    from the attention how to make this long

  104. 4:11

    context.

  105. 4:12

    >> Yeah, I would say the story about long

  106. 4:14

    context went back to even Minimax M1 and

  107. 4:17

    Minimax 01 where the model was actually

  108. 4:21

    was able to perform tasks of 10 million

  109. 4:25

    token context.

  110. 4:26

    >> 10 million?

  111. 4:26

    >> 10 million, yes.

  112. 4:28

    But then it was not an agentic model,

  113. 4:31

    right? It was just 10 for example, you

  114. 4:33

    dump in a book, it would be able to give

  115. 4:35

    reviews on it and stuff like that. So

  116. 4:38

    what we realized was that, you know,

  117. 4:41

    longer context actually unlocks a lot of

  118. 4:43

    capabilities especially when interacting

  119. 4:47

    with users and now when, you know, the

  120. 4:49

    agent is interacting with the whole

  121. 4:51

    environment and getting all the tool

  122. 4:53

    responses,

  123. 4:54

    getting multi rounds

  124. 4:57

    the like shorter context wouldn't be

  125. 5:00

    enough to perform the complex task. So

  126. 5:04

    for this version we said, "Oh, we have

  127. 5:07

    to have our longer context backs."

  128. 5:10

    So what we pursued was with our Minimax

  129. 5:14

    sparse attention.

  130. 5:15

    Um which you know, was the architecture

  131. 5:20

    that was scalable

  132. 5:22

    and had a simple design. So I would say

  133. 5:26

    from a higher level, right? It has an

  134. 5:28

    index branch

  135. 5:30

    that, you know, selects

  136. 5:33

    on a higher level what is what matters

  137. 5:35

    more in the context and then we have a

  138. 5:38

    sparse attention branch that calculates

  139. 5:40

    performs the calculation on the selected

  140. 5:42

    blocks to actually performs the task.

  141. 5:45

    Um and so

  142. 5:47

    yeah, like that we really designed um an

  143. 5:51

    elegant architecture so that we can

  144. 5:53

    scale the length and then scale the

  145. 5:55

    model size in the future with that.

  146. 5:57

    >> That's beautiful. I like how for those

  147. 5:59

    who've been in the field for quite some

  148. 6:01

    time, we we had a lot of work on

  149. 6:02

    attention, right? This N square and

  150. 6:04

    there was a lot of linear attention.

  151. 6:06

    >> Yeah.

  152. 6:06

    >> And then some that somehow all of these

  153. 6:08

    disappeared at some point when flash

  154. 6:10

    attention came around. We discovered we

  155. 6:11

    just needed more efficient camera. Now,

  156. 6:13

    I like how we come back to thinking, you

  157. 6:15

    know,

  158. 6:16

    first principle, what is attention? How

  159. 6:18

    can we make that more efficient? So, 1

  160. 6:21

    million token is crazy, right? GPT-2 was

  161. 6:24

    1,024 and and everyone was like, "Oh,

  162. 6:26

    that's really big. We we we never need

  163. 6:28

    more."

  164. 6:29

    Where do you see this coming? Like,

  165. 6:30

    going in the future? Like, Jeff Dean was

  166. 6:32

    pitching me the other day a trillion

  167. 6:33

    token attention. You think we should go

  168. 6:36

    to

  169. 6:36

    a trillion token attention?

  170. 6:38

    >> So, that's definitely something we can

  171. 6:39

    explore towards, right? Ultra lengths of

  172. 6:43

    the context. Definitely, it's something

  173. 6:46

    that's very exciting to explore with.

  174. 6:47

    And something that architecture design

  175. 6:49

    along with hardware um

  176. 6:52

    would require a lot of research onto

  177. 6:53

    that. Yeah.

  178. 6:54

    >> You think there's still a lot of

  179. 6:55

    low-hanging fruits? So, typically today

  180. 6:58

    we saw Open AI really reducing I mean,

  181. 7:00

    we don't know how as of firm, but like

  182. 7:02

    reducing their their inference bill by

  183. 7:04

    half, by probably having some more

  184. 7:06

    efficient processing around tensions of

  185. 7:08

    any type. You think there's still a lot

  186. 7:10

    of low-hanging fruit that can be getting

  187. 7:12

    um

  188. 7:13

    how we can process that. So, so one one

  189. 7:15

    thing we're still very interesting about

  190. 7:16

    M3 is how cheap it is in particular

  191. 7:18

    because of this part attention and part

  192. 7:20

    because of its small one, but it's also

  193. 7:22

    very efficient.

  194. 7:23

    >> Right.

  195. 7:23

    >> You think we can go even way further?

  196. 7:25

    And maybe how did you guys invented uh

  197. 7:29

    Min Max Fast Attention? Was it an agent

  198. 7:31

    coming up with the idea? Was it a human

  199. 7:32

    still coming up with the idea? Tell us a

  200. 7:34

    little bit about

  201. 7:35

    >> Yeah. Um, so we do think there's still a

  202. 7:39

    lot of work that can get into

  203. 7:41

    architecture and inference optimization

  204. 7:43

    so that the model can be more efficient,

  205. 7:46

    especially if there are tasks that are

  206. 7:48

    very task sensitive but require very

  207. 7:51

    strong capabilities, right? And for that

  208. 7:53

    those kind of tasks we really want model

  209. 7:55

    to be efficient.

  210. 7:57

    Um and who came up with this part of

  211. 7:59

    actually I think an intern from our team

  212. 8:02

    worked on that. That yeah, an intern.

  213. 8:05

    Uh that

  214. 8:06

    doesn't usually happen in a lot of labs

  215. 8:09

    because I think in some labs interns

  216. 8:11

    don't have access to

  217. 8:13

    the data, the work, and stuff. Uh but

  218. 8:16

    yeah, we are open to anyone who would

  219. 8:19

    like to contribute to our models. So, um

  220. 8:21

    the architecture was actually designed

  221. 8:23

    by an intern.

  222. 8:24

    >> It's very good. Still some work for

  223. 8:25

    interns here. Good good news. Um that's

  224. 8:29

    also a good segue to also how Min Max is

  225. 8:31

    working internally. So, so we were

  226. 8:33

    discussing before coming on stage I was

  227. 8:35

    saying everyone can propose a project.

  228. 8:38

    Can you tell us a little bit about how

  229. 8:39

    you are organized, how you do your

  230. 8:41

    research?

  231. 8:42

    >> Mhm. Mhm. I think that is very different

  232. 8:44

    from uh

  233. 8:46

    even in school or even in earlier

  234. 8:50

    you know, the earlier tech companies is

  235. 8:52

    pretty pretty different. It's that um

  236. 8:55

    what we

  237. 8:56

    what we make sure is that we have good

  238. 8:59

    foundation and good um infrastructure so

  239. 9:02

    that anyone can play with the model and

  240. 9:05

    can think of what they can improve with

  241. 9:07

    the model.

  242. 9:08

    And then after model releases when they

  243. 9:10

    are free, right? They can play with the

  244. 9:13

    model. They can think of their own

  245. 9:15

    evaluations. They can find their own

  246. 9:17

    weaknesses and propose a thing that they

  247. 9:19

    want to improve on the model. And then

  248. 9:22

    other people who are interested in that

  249. 9:24

    would, you know, propose to join the

  250. 9:25

    project and they will work on for a

  251. 9:27

    couple of weeks or even a couple of

  252. 9:29

    months. And when they work out, the

  253. 9:31

    final thing is shipped to our model. It

  254. 9:34

    it is, you you we use that in our final

  255. 9:36

    training and it's shipped out to the

  256. 9:38

    audience.

  257. 9:39

    >> Interesting. So, you can have people

  258. 9:40

    working for a really long time on

  259. 9:42

    project. When you say a couple of

  260. 9:43

    months, it can be like really deep

  261. 9:45

    exploration of

  262. 9:46

    >> Yes.

  263. 9:46

    Yes.

  264. 9:48

    I would say, for example, architecture

  265. 9:50

    might require longer time of

  266. 9:51

    investigation, research, experiments,

  267. 9:54

    even redoing the evaluations for

  268. 9:56

    pre-training.

  269. 9:58

    Yes, so it might require longer time.

  270. 10:00

    >> Very nice, yeah. And I know you're also

  271. 10:02

    very big on evaluation. I agree. We

  272. 10:03

    could talk about that. I think what One

  273. 10:05

    thing probably related to that is this

  274. 10:07

    unique specificity that M3 and your team

  275. 10:09

    has um around multimodality. So, not

  276. 10:12

    just text, but the similar can also

  277. 10:14

    understand image and video. And as I

  278. 10:17

    understand, but but please explain

  279. 10:18

    explain better,

  280. 10:20

    when you read the model card on Hugging

  281. 10:21

    Face, it say the model was trained from

  282. 10:23

    the first step as a multimodal, not just

  283. 10:25

    have a like user one as after solved,

  284. 10:27

    right? Can you tell us a little bit more

  285. 10:29

    about that and why you think it's

  286. 10:31

    important and

  287. 10:32

    and and why starting from the first step

  288. 10:34

    on multimodal training and not just just

  289. 10:36

    training this.

  290. 10:37

    >> Um so, we call it native multimodality.

  291. 10:40

    Um and so, it is somehow typical for

  292. 10:45

    model labs to train the multimodal,

  293. 10:48

    let's say, vision understanding

  294. 10:49

    capabilities after the text pre-training

  295. 10:52

    is done. Um they put adapters and then

  296. 10:55

    train that part.

  297. 10:57

    But what we found out was that that

  298. 10:59

    would actually harm the text

  299. 11:01

    performance.

  300. 11:03

    And the vision vision understanding

  301. 11:05

    performance wouldn't converge that well

  302. 11:07

    because the model is kind of converges

  303. 11:09

    towards the model the text

  304. 11:10

    understanding. Um and it's just not the

  305. 11:14

    most optimal. And also not the most

  306. 11:17

    scalable, if you think about it. We want

  307. 11:18

    to scale the data, right?

  308. 11:20

    And also, we can also some labs um train

  309. 11:24

    this capability from halfway through the

  310. 11:27

    pre-training. For example, continued

  311. 11:29

    pre-training. But what we found that

  312. 11:31

    this would be very, you know, uh recipe

  313. 11:35

    sensitive.

  314. 11:36

    It is different for the recipe would be

  315. 11:38

    different for different architectures,

  316. 11:40

    different, you know, data mixtures,

  317. 11:42

    different learning rates.

  318. 11:44

    It's hard to control, hard to, you know,

  319. 11:46

    scale to

  320. 11:49

    you can't really scale your experiment

  321. 11:52

    results and conclusions to a larger

  322. 11:54

    model.

  323. 11:55

    And so

  324. 11:57

    you know, what we thought was why not

  325. 11:59

    just training from the very first step?

  326. 12:01

    That comes to the most natural. We know

  327. 12:04

    that a lot of labs run into problems

  328. 12:06

    doing that.

  329. 12:08

    The model would collapse after a couple

  330. 12:10

    of steps of training, you know, both

  331. 12:13

    text and vision understanding, but we

  332. 12:16

    managed to solve that problem.

  333. 12:18

    We did a lot of work on

  334. 12:21

    VIT and we did a lot of work on the data

  335. 12:24

    that we actually training. For example,

  336. 12:27

    we do interleave the data, what we call

  337. 12:29

    interleave the data.

  338. 12:31

    It's actually natural data, but we keep

  339. 12:33

    the

  340. 12:35

    images and videos in instead of masking

  341. 12:37

    it out and we do some pretty

  342. 12:40

    good cleaning and masking on the data

  343. 12:43

    and we do very good reward modeling so

  344. 12:45

    that we train it from the first step and

  345. 12:47

    scales up

  346. 12:50

    a lot. Yeah, it does does not collapse.

  347. 12:53

    >> That's really impressive. Impressive.

  348. 12:55

    Should we Should we expect much larger

  349. 12:57

    model in the future? So this one is

  350. 12:58

    still fairly small, right?

  351. 13:00

    >> It's It's 428 billion parameters, 23

  352. 13:03

    active billion.

  353. 13:05

    Well, do you think you will go past the

  354. 13:07

    trillion?

  355. 13:09

    >> Definitely. Yeah, definitely in the

  356. 13:11

    future.

  357. 13:13

    There are many tasks that wouldn't be

  358. 13:15

    able to the more model

  359. 13:17

    wouldn't be able to perform very good at

  360. 13:19

    with smaller parameters. We are

  361. 13:22

    definitely going more ambitious than

  362. 13:24

    this.

  363. 13:25

    >> It's great. Looking forward. Um another

  364. 13:28

    interesting thing I

  365. 13:30

    I always been find fascinating about Min

  366. 13:32

    Max is how how you also have this whole

  367. 13:34

    range of of apps and product, right? So,

  368. 13:36

    I remember already So, so Min Max

  369. 13:39

    started to open source things on the on

  370. 13:41

    the hugging face platform in in January

  371. 13:43

    last year. So, that 18 month ago and and

  372. 13:45

    we were chatting a little bit about the

  373. 13:47

    team to understand what you were doing

  374. 13:48

    and I remember you So, you were already

  375. 13:50

    having a huge usage on some of these of

  376. 13:53

    some of these apps. Um can you tell us a

  377. 13:56

    little bit how how this started, right?

  378. 13:58

    So, was it basically you had a lot of

  379. 14:00

    apps and then you thought we have we

  380. 14:01

    have all this data, why not training a

  381. 14:03

    model and then they build up research

  382. 14:05

    team. How is how is the story there?

  383. 14:07

    >> Um our our story is modeled from the

  384. 14:11

    first day. So, um I believe that

  385. 14:14

    multi-modality model a model that can

  386. 14:16

    understand all visions and outputs all

  387. 14:19

    modalities was the first thing that our

  388. 14:22

    um CEO planned on the first day even

  389. 14:24

    before the company even started. So,

  390. 14:26

    that was the dream of AGI. I think that

  391. 14:28

    was very very early even before ChatGPT

  392. 14:30

    came out.

  393. 14:31

    >> Wow.

  394. 14:31

    >> Um yeah.

  395. 14:33

    And then apps were something that comes

  396. 14:35

    along because you have some model

  397. 14:37

    capabilities you want people to

  398. 14:38

    experience it well.

  399. 14:40

    Not many people can use it with API,

  400. 14:43

    right? We can't expect everyone to

  401. 14:44

    experience with API. So, we need good um

  402. 14:48

    user interaction,

  403. 14:50

    you know, interfaces, good apps, good

  404. 14:52

    scenarios that people can

  405. 14:55

    can you know, experience experience

  406. 14:57

    model with. I think actually those apps

  407. 15:00

    covered more than 300 million people

  408. 15:03

    around 200 countries globally.

  409. 15:07

    And I think over a million companies as

  410. 15:09

    well.

  411. 15:10

    >> Yeah, this was a mind-blowing when I

  412. 15:11

    heard about the the size and we we don't

  413. 15:13

    often realize the size of of of this

  414. 15:16

    type of usage already. And and that I

  415. 15:18

    kind of brings me to the question around

  416. 15:20

    um

  417. 15:22

    open source business model, and and all

  418. 15:24

    of that, which is the

  419. 15:26

    always existing question, which is right

  420. 15:28

    now it's nice to open source model, but

  421. 15:30

    you you also need to have some revenue

  422. 15:32

    stream, right?

  423. 15:34

    So, I guess M3 is something you you

  424. 15:36

    decided, for instance, to be for free,

  425. 15:38

    and I'm I think it's it's it's great for

  426. 15:40

    the world. Um how do you see this? Do

  427. 15:43

    you also have some specific models you

  428. 15:44

    use for the app? Do you think about Do

  429. 15:47

    you think in the future you'll keep It's

  430. 15:49

    probably hard to say for sure, but do

  431. 15:50

    you think you'll keep open sourcing

  432. 15:52

    models? How is the culture around open

  433. 15:54

    sourcing right now?

  434. 15:55

    >> Personally, and also for the model

  435. 15:57

    research team, we always hope to open

  436. 16:00

    source the models. Um that is our plan.

  437. 16:03

    Because we

  438. 16:04

    really see how the open source community

  439. 16:06

    together can help the model build

  440. 16:09

    better. For example, we receive a lot of

  441. 16:12

    um feedbacks on the model performance

  442. 16:14

    from the great community, and we receive

  443. 16:16

    PRs on uh wh- whatever we open source.

  444. 16:20

    And those are very very valuable and

  445. 16:22

    come comes to our later versions. So,

  446. 16:24

    definitely open sourcing is great.

  447. 16:27

    >> That's great. And actually, do you have

  448. 16:29

    some ask for the audience, people who

  449. 16:31

    are using uh M3 or MiniMax? Is there

  450. 16:33

    something you would love them to send

  451. 16:36

    back to you as feedback? Do you Do you,

  452. 16:38

    for instance, do you read when people

  453. 16:40

    try to modify the models or play around,

  454. 16:42

    you know, tweaks? Or what is the best

  455. 16:45

    thing you you think you can take from

  456. 16:47

    the community uh for future models, for

  457. 16:50

    instance?

  458. 16:50

    >> Mhm.

  459. 16:51

    I would say whatever um issues that

  460. 16:54

    people are running into, especially with

  461. 16:56

    multimodality, right? This is the first

  462. 16:58

    time that we're combining it together.

  463. 17:00

    We are definitely going more ambitious

  464. 17:02

    on that in the future. It might have

  465. 17:04

    some flaws right now, but we are

  466. 17:05

    improving on that. So, whatever that's

  467. 17:08

    uh feedback that model is not good doing

  468. 17:10

    that great, we will definitely improve

  469. 17:12

    that in future versions. And also,

  470. 17:14

    whatever features that uh people want.

  471. 17:17

    Say,

  472. 17:18

    you know, for example,

  473. 17:20

    thinking effort. Right? Some people ask

  474. 17:22

    for that. Um

  475. 17:24

    like everyone can ask, and we will try

  476. 17:26

    to accomplish that in the future models.

  477. 17:29

    >> Yeah. Do you see a lot of usage right

  478. 17:30

    now already in multimodality in terms of

  479. 17:33

    coding agents? I feel like it's it's a

  480. 17:35

    little bit un- underexplored.

  481. 17:38

    >> It is It is, but um it can actually

  482. 17:42

    unlocks a lot of capabilities and a lot

  483. 17:45

    of

  484. 17:45

    uh agent applications.

  485. 17:47

    Say that for example, you want a model

  486. 17:49

    to read a PP- PowerPoint, uh or to read

  487. 17:52

    some report that is not very structured.

  488. 17:55

    Um and you want it to understand a very

  489. 17:58

    long video, say that you dump in a

  490. 18:00

    long-playing video, and then you want

  491. 18:02

    the model to act uh using some tools uh

  492. 18:05

    after understanding it. And it unlocks a

  493. 18:08

    wide variety of um agent use cases.

  494. 18:12

    >> So, like the agent could finally watch

  495. 18:14

    my YouTube tutorial and understand how

  496. 18:16

    to use my coding tools, how I described

  497. 18:18

    it? Is it something like that?

  498. 18:20

    >> Uh

  499. 18:21

    >> Could the agent finally watch YouTube

  500. 18:23

    tutorials and understand things from

  501. 18:24

    them?

  502. 18:25

    >> Yeah, yeah, yeah.

  503. 18:26

    I think so.

  504. 18:27

    >> Do you use a lot of uh agent coding

  505. 18:29

    tools internally? Is it like I mean,

  506. 18:31

    coding for sure, but like is it also

  507. 18:33

    already in terms of research? Is it

  508. 18:35

    automated part or not? How How does this

  509. 18:38

    uh

  510. 18:38

    >> Yes.

  511. 18:39

    >> work?

  512. 18:39

    >> Um we have our own research harnesses.

  513. 18:42

    Um we build our own research harnesses

  514. 18:44

    that automate our workflows. I would say

  515. 18:47

    a lot of our workflows are automated.

  516. 18:49

    You can see how um the latest frontier

  517. 18:52

    models all pursues capability less

  518. 18:56

    kernel optimization, right? Like let the

  519. 18:58

    model post string other models.

  520. 19:01

    Um let the model build data.

  521. 19:03

    Auto data, stuff like that. Um you can

  522. 19:06

    see how more and more models are capable

  523. 19:09

    of doing those, including M3. Actually,

  524. 19:11

    we were very good at those cases, longer

  525. 19:14

    horizons and coronal organizations.

  526. 19:16

    Um and so we can use that model

  527. 19:19

    capability, harness it together, and

  528. 19:22

    help with our um daily routine, and make

  529. 19:25

    our iterations even faster.

  530. 19:27

    >> Is M3 building M4 already?

  531. 19:29

    >> Um building M3.1.

  532. 19:32

    >> M3.1, okay.

  533. 19:33

    >> [laughter]

  534. 19:34

    >> Let's hit the gym.

  535. 19:34

    >> Already.

  536. 19:36

    >> Um I I would love to finish on what what

  537. 19:38

    you find exciting in the coming month,

  538. 19:41

    what what do you think it it can be Asia

  539. 19:43

    in terms of feature or or things you

  540. 19:45

    want to see happening in in AI or more

  541. 19:48

    generally in terms of whatever whatever

  542. 19:51

    really is top of your mind I would say

  543. 19:52

    and it's going to

  544. 19:54

    happen.

  545. 19:55

    >> A lot of things are very exciting. Um

  546. 19:58

    but what I recently find the most

  547. 20:00

    exciting would be a multi-agents that I

  548. 20:03

    think a lot of AI applications are

  549. 20:05

    using, model routing, multi-agents um

  550. 20:09

    that allows even more capabilities, even

  551. 20:12

    more complex tasks, and also it tells us

  552. 20:16

    what the models are capable and not

  553. 20:18

    capable of, and you can, you know,

  554. 20:21

    do a lot of things with that. It's

  555. 20:23

    pretty exciting.

  556. 20:24

    >> Thanks a lot, Alif. Pleasure to have

  557. 20:26

    you.

  558. 20:27

    >> Thanks for having me.

  559. 20:28

    >> Thanks, everyone.

  560. 20:30

    >> [applause]