Artificial Intelligence
Joined March 2021
- Tweets1.1K
- Following538
- Followers8.5K
- Likes39.6K
The Kibitz model claims to be a Kimi model. The model also currently appears at a very high rate in the Code Arena
Asking models about their identity in the Code Arena is unreliable. GLM and other models also sometimes identify as Kimi models, so I’m not sure yet what the Kibitz model is
There are new rumors that the next GLM Model, GLM 5.5, will have 2.5T - 3.5T parameters and significant architecture changes
In its August HKEX interim filing, Zhipu AI said it had validated the idea of Looped Transformers in preliminary experiments. I wonder if we'll see it implemented in the next GLM model
🤖 Made with AI
关于 @Zai_org 下一代GLM5.5的一些最新消息：
- 下一代基座规模在2.5万亿至3.5万亿之间，比此前流传的1.6T大了不少。架构改动也比较激进，希望同时提高训练和推理速度，把推理成本压下来，目标是在发布第一天就能提供大规模服务。
- 前段时间基座训练出现了loss不降的情况，进度因此慢于市场预期。这个问题近期应该已经解决，预计能较快完成训练。
- 预计下一代GLM模型能达到Fable级别的旗舰水平。
- 会继续朝Claude的使用体感推进，在后训练阶段大量投入任务环境，通过持续迭代提升模型能力。GLM-5系列的一个基座可以使用半年以上，在同一基座上持续做后训练、更新能力，这也是他们目前比较明确的技术路线。
- 之前的RSI已经取得了一些进展，但目前还没有成为模型能力提升的主力，预计会在未来某个版本中进入服务。
我个人视角比较期待的是，这次GLM-5.5能不能把3T左右的模型规模、旗舰级的能力、推理速度和调用成本一起做好。
尤其是拿模型跑Agent，任务一长，就要连续调用很多轮。模型能不能把事情做完、要等多久、最后花多少钱、服务会不会限流，都会直接影响实际使用和销售的情况。
Kimi K3.1 was recently spotted in 2 SGLang commits, one of which references a K3.1 deployment already running on Deepinfra
If we are getting K3.1 on the Arena this week, it might follow the same naming pattern as the previous Kimi models:
Kimi K2.5 - Kiwido
Kimi K2.6 - Kiwire
Kimi K3 - Kivine
Kimi K3.1 - ???
The FAQ on the MinMax site claims that 3.1 Flash has the same paramters as MinMax M3. 428B / 23A
Source:
agent.minimax.io/tools/m3-1-…
Some new Claude EAP model names have been spotted in the Devin CLI binary
"Verrine" and "Onigiri" seem to have been added about a week ago
Reliabe leakers have said that partners got access to Fable 5.5 recently. A release next week seems likely
