GLM-5.2: Open Weights, Near-Frontier Intelligence — Zixuan Li, Z.ai

AI Engineer · 13 min · 139 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

  1. 00:12Without further ado, we're going to bring on Zishan for for his keynote if we're all set.
  2. 00:20Hey, how are you?
  3. 00:22Great I'm here.
  4. 00:23Yes, to the magic of the internet.
  5. 00:26It's really good to see you again.
  6. 00:27Zishan has you've been speaking with AI Singapore and obviously GLM is the talk of the town right now.
  7. 00:35There's been an amazing booth downstairs in the expo and I think a lot of people are just very excited to hear from you on what all is going on with ZIAI.
  8. 00:43So if you want to take it away.
  9. 00:47Yes, sorry again for not showing up in person, but I have a team whole team coming to the town and we have a booth.
  10. 00:55So you have any questions, feel free to reach out to me on X link and also you can look for my team on our booth.
  11. 01:05And since I cannot see my slides, so I will need Swee to help me flip through all the slides for me.
  12. 01:11Yeah.
  13. 01:14You're good.
  14. 01:14Here Swee, like what which slides we are in right now?
  15. 01:17Yeah, so we're we're on the opening slide.
  16. 01:20We're talking about intelligence and ZIAI.
  17. 01:25Okay, so yeah, so you can you can see that it's the first time for us to introduce GLM 4.2 5.2 to the world and also we are going to share something
  18. 01:35about Z.AI and GLM because maybe people will think Z.AI and GLM they're irrelevant, right?
  19. 01:43You are your company is not G.AI or your model's called Z1 or Z2.
  20. 01:49And you can find my X and our company's X account here, so you can just search for my name Zishan Lee and the Z A I org.
  21. 01:58And you can follow them for the follow-ups.
  22. 02:01So, yeah, we can go to the second slide.
  23. 02:05Yeah, actually I cannot see the slide, so I'll try to
  24. 02:12So, we we
  25. 02:12Yeah, try to make sure that it is correct.
  26. 02:15So, the company actually is called Zhipu.
  27. 02:19Maybe some people have heard of it.
  28. 02:23And the all the models, it's it's called GLM.
  29. 02:27Actually, it's not a a brand name, it's a generic term.
  30. 02:32So, GLM actually represent general language model pre-training with auto-regressive blank filling.
  31. 02:38And that paper was published back in 2021.
  32. 02:42So, actually we were the one of the first labs to do explorations on large language models at the same time with OpenAI, and Anthropic, and DeepMind.
  33. 02:53And even today, we we no longer use GLM as the architecture, we still use the name GLM as our brand name.
  34. 03:00So, we use GLM 5.1, 5.2, and it become like one of our like most prod uh proddest product and model.
  35. 03:11And the second thing that we look for is intelligence upper bound.
  36. 03:16So, in terms of intelligence, we we may feel that it's represent IQ, something something like that.
  37. 03:23And when Deep Seek launched one 01 launch, people are talking about the model's capability to solve math problems, physics problems.
  38. 03:32But what actually intelligence mean is not just like IQ or AIME or other other physics problem.
  39. 03:40So, from GLM 4.5 to GLM 4.7, we are exploring like several things, like reasoning, coding, and agentic capabilities.
  40. 03:49So, as you can see from the slides, So, we add like Yeah, the last slide.
  41. 03:55Yeah, we were actually I haven't like finished that slide.
  42. 04:00Yeah, so Okay.
  43. 04:03Yeah, need to go back to the two two slides.
  44. 04:08I don't have back I don't have a back button.
  45. 04:15Thank you.
  46. 04:17I would you I don't understand clickers that don't have back buttons.
  47. 04:19Like why Okay, I you know, anyway, go ahead.
  48. 04:24Yeah, never mind.
  49. 04:25Yeah, because people want to see the GLM 5.2. They they don't want to see you like GLM 4.1 or 5. But like GLM 4.2 actually specialize in coding and genetic task as you can see from the graph because
  50. 04:38there are a lot of rumors whether your model is close to mythos, fable, but actually I want to share these slides to all of you.
  51. 04:45So, you can see it's somewhere between Opus 4.7 and 4.8 and we use the hardest problems like Deep Sweep, Terminal Bench 211, which was mentioned by the Open AI team
  52. 04:58uh several minutes ago.
  53. 05:01And all the like long horizon task and benchmark shows that the capabilities are on par with at least Opus 4.7 and it it it shows a significant improvements over
  54. 05:161.1. Also, for GLM 1.2, we add a thinking uh level called high.
  55. 05:23So, because we also noticed as we move to the harder task it may consume more tokens and also we we care a lot about the token efficiency.
  56. 05:34So, it's the first time we add the high level for thinking budget.
  57. 05:38But even without thinking, the non-thinking model is better than the 5.1 thinking model.
  58. 05:45So, I think it's a huge improvement for the open weight model.
  59. 05:49That's what really impressed the world and why people are talking about GLM 5.2 lately.
  60. 05:56Okay, the next slide.
  61. 05:59And one one thing that I want to mention is that GLM is more more than coding model because people use it inside clock code code X, open code.
  62. 06:09But actually, we have trained a lot of things outside coding.
  63. 06:13For example, we improve a lot in GDP valve and also math problems.
  64. 06:20We also care about math problems, frankly speaking.
  65. 06:23And also, we train a lot of thing uh related to role play, general chat.
  66. 06:28We want to improve every aspect of the model.
  67. 06:32So, you can see from the artificial analysis intelligence index actually leads the other open way model a lot and close to the frontier model.
  68. 06:42So, I want you if you are you haven't experienced GLM yet, you can use GLM to do general chat, use it to process your daily workflow, not just for coding, but you can explore the model
  69. 06:56like beyond the coding scope.
  70. 06:59Next slide.
  71. 07:01And GLM 4.2 is a open weight model.
  72. 07:06So, people always ask me, "Why do you open weight?
  73. 07:09So, do you care about your business or like do you care about losing market to some infra suppliers?"
  74. 07:17But actually, we open the weights for several things because there are users' needs and there are our own needs.
  75. 07:24If we can meet their needs, I think it's okay.
  76. 07:28It's definitely okay for us to to open the model.
  77. 07:31For example, if our users want security and control and we want to build trust, we can open weight model.
  78. 07:40For for some cases, if enterprise or government especially in the Western world, want to use the model, we open way, we upload to the hugging face so that they can
  79. 07:51use the model on premise.
  80. 07:53I think it's very beneficial for the whole ecosystem it to explore the model, especially when the capabilities is close to the frontier model.
  81. 08:03And second, if people want diversity.
  82. 08:07So, in terms of diversity, I mean the capabilities in legal, finance, security, they can fine-tune the model.
  83. 08:15So, we need to open the model to let them fine-tune.
  84. 08:19For example, Harvey is fine-tuning GLM 4.1, and maybe they're thinking about fine-tuning GLM 5.2 afterwards.
  85. 08:29And I've I have like talked to a lot of other companies.
  86. 08:33They're also thinking about fine-tuning GLM as their like next step or next strategy to differentiate themselves from other application company.
  87. 08:45And third, if our customer or an individual want to co-design and predict the future, sometimes they need to see the architecture of the model.
  88. 08:57They need to see the recipe of how you train the model.
  89. 09:01So, we want to make the norm.
  90. 09:05We want to make my bets.
  91. 09:07So, we want to co-shape the future with our customers with the open source uh community.
  92. 09:13So, I think our needs and our ecosystem really like fits into each other.
  93. 09:20And GLM 5.2 couldn't succeed without you, all the open source community players like Onflos and Media and and some like individuals, super developers, they are part of it.
  94. 09:35And thanks a lot to application builders like Peter.
  95. 09:40Right?
  96. 09:40Because open source doesn't just include open source model, but also open source softwares and open source other sorts of support.
  97. 09:49So, all your support and what are you you're doing right now pushes to to make better models.
  98. 09:59I think uh you you are the true hero.
  99. 10:02And the last, I want to share a a great resource for you to go through.
  100. 10:08Geo 5.2 is our tech blog.
  101. 10:11So, actually in that tech blog, we share several things like our hugging face repo, how how you can try the Geo and 5.2. You can try the inside the chatbot
  102. 10:24agent, and also you can call the API.
  103. 10:27And also we have a coding plan like the Codex or cloud core subscription for you to use your tokens as individuals.
  104. 10:36And also we share something about our training pipeline training recipe, which you can understand why it's a good model.
  105. 10:43So, there are a lot of things behind the behind the model, not just um a a model that had great data.
  106. 10:51We also have fantastic technologies behind the model, so you can export the model yourself.
  107. 10:57You can see what difficulties we have gone through.
  108. 11:02And the last slide, actually it's kind of a one more thing.
  109. 11:07So, it's the first time we share the code to the whole community.
  110. 11:13Woo!
  111. 11:14So, that one more thing is we actually have our own harness.
  112. 11:19The Z code, actually it's built for Geo 5.2, but also support all frontier models.
  113. 11:26You can bring your own key.
  114. 11:27You can connect to Z code.
  115. 11:29Actually, the I think the operation is is similar to Kodak.
  116. 11:34Actually, you you can try some techniques like go or other like compact technique techniques like what you've done in in Kodak and car code.
  117. 11:46And this harness is I think it's the perfect one for GLM 102.
  118. 11:51If you haven't experienced it, you can just search Z code or you can go to our booth.
  119. 11:57We have our team members showing this harness to you.
  120. 12:01And welcome to to our booth.
  121. 12:04And welcome to like talk to me in the future.
  122. 12:07And next time I'll definitely be in SF talking to everyone.
  123. 12:11Yeah, thanks.
  124. 12:13Thank you very much.
  125. 12:17Um So uh trust me, the very first World's Fair, I wasn't allowed to come back in the country for my own conference, so I know exactly this feeling.
  126. 12:29Uh but it it it's all good.
  127. 12:31Uh the ZAI team, I really appreciate them uh making the effort.
  128. 12:34They really want to meet you.
  129. 12:35They're here to meet you.
  130. 12:36And this is the whole point of the World's Fair, to bring all the top world's AI uh companies and labs uh all in one place, so you can do business together,
  131. 12:45uh meet the people behind the models that you use, ask the questions that he cannot answer in public, but you can ask in private.
  132. 12:52Um I'm also very proud.
  133. 12:54Uh he showed uh Zhicheng showed that that list of uh Hugging Face uh you know, top uh contributors.
  134. 13:00I think we are four for eight uh in that list uh present at World's Fair.
  135. 13:04Uh we're working with Nvidia and Unsloth and Ollama and then uh all those other fine-tuners as well to uh to make our our sort of the inference and the local tracks uh that you're going to see over the next few days.
  136. 13:16Um with that, thank you so much to Zhicheng.
  137. 13:18Uh I'm going to invite on uh the next uh couple folks.
  138. 13:21I think I I I think I might be doing Ali's job here, so uh we'll we'll talk to Hugging Face next and Minimax.
  139. 13:28Thank you.