One more detail is that this release points to a discontinuous maturation in Zhipu's post-training stack. Yeah distilled from Claude, whatever, it served as a seed dataset. But you don't get those CritPT/Posttrain-Bench/WeirdML etc. results just by finetuning on claudisms. They now know how to build agents which synthesize their own diverse training environments and get stronger. They are entering the early stage of RSI. The ceiling of this approach is very far away, they can have steadily improving GLM-5s every 2 months until the end of the year without doing anything new. I'm pretty certain that the next one will be stronger across the board than Opus 4.8 (maybe modulo some holdovers like WeirdML). The gap may be stable, or modestly decrease, or increase if you count Fable/Mythos, but there is no slowdown in open weights capabilities progress, and here we clearly have something that's *at least* on "Opus 4.55" or "GPT 5.4" level. Opus 4.5 was already a paradigm shift (and I'd argue that GPT 5.2 was a bigger one). We have that on huggingface now. It can help build more of itselves. Make of that what you will.
Zvi on GLM 5.2. Mostly correct. One detail he underrates (again) is that model layer and hardware layer are distinct. 5.2's niche is, ironically, larger than what http://Z.AI can serve as a product. On B300s, we could run it *faster* and cheaper than Gemini-Flash.
