Zhipu's Jie Tang: GLM-5.3 Matches GLM-5.2 in Every Parameter, Gains Come From Post-Training Alone
AccBalanced · x · 2026-08-20
Tsinghua professor and Zhipu founder Jie Tang argued on X that total parameters matter only up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass and, above all, post-training.
He explicitly framed GLM-5.3 as the controlled experiment for that claim: same base, same architecture, and same total and activated parameter count as GLM-5.2, with one month of scaling long-horizon environments and RL — and the gains are "not marginal."
In the quoted long post, Tang argues parameter count is only meaningful alongside how much data you have, where compute will be spent, and who runs the model under what conditions. He recounts the field's hard lessons: Kaplan et al. (2020) fit an exponent telling everyone to grow parameters faster than data (2.7:1), which the industry followed with GPT-3, Gopher and MT-NLG; Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter. His conclusion: scaling has more than one dial — this time Zhipu turned the post-training one.
More from Models
- Meta's coding-optimized open-weight Muse Spark Open lands on Modal with 1M context — AAAzzam · 2026-10-16
- Google launches Gemini 4 Argon frontier model on Cloud with 10 official demos — Saboo_Shubham_ · 2026-10-16
- RL practitioner: ~10,000 tasks still stump every frontier coding model — max_paperclips · 2026-10-16
- SpecID eval falls: Gemini 3.5 Flash scores 67%, beating its human creator at Zimbabwean species ID — johnowhitaker · 2026-10-16
- Researcher criticizes Claude for lecturing users and refusing critical takes on tech firms — BlancheMinerva · 2026-10-16
- Chart: Meta went 492 days without an open-weight release before Muse Spark Open — demian_ai · 2026-10-16