One interesting thing that makes @jietang's prediction believable is that GLM has an unusual pace in base model iterations, which is a bit obscured by versioning. GLM-4, 4.1, 4.5 are all different pretrains. I can see them milking 5th gen to Q4 2026 and making a jump.
“Won’t take that long” ~ Jietang
For GLM 5.2 to reach the Fable tier.
No Digg Deeper questions have been answered for this story yet.
In contrast, Kimi K2 is a year old and K2.7-Code is a direct descendant. DS-V3 series lasted for a year too. Some Opuses, Sonnets, GPTs keep working for over a year. Qwen is faster, but Alibaba is a vastly larger entity.
One interesting thing that makes @jietang's prediction believable is that GLM has an unusual pace in base model iterations, which is a bit obscured by versioning. GLM-4, 4.1, 4.5 are all different pretrains. I can see them milking 5th gen to Q4 2026 and making a jump.
Of course, it's one thing to tinker with 32B llama-likes, then moving onto a 355B MoE, then to 744B DeepSeek clone, and quite another thing to train anything close to Mythos, which OpenAI has so far not succeeded at. I'm skeptical.