ZhiPu GLM-5.3 Boosts Coding 50% via Post-Training, No Base Model Change
ZhiPu co-founder Tang Jie shared that GLM-5.3, based on GLM-5.2 (743B MoE, ~40B active params), improved coding ability by 50% through one month of reinforcement learning in long-cycle environments. The post distinguishes total parameters (knowledge) from activated parameters/effective depth (reasoning), noting that security vulnerability finding relies on reasoning chains, not databases. It outlines scaling history: Kaplan et al. (2020) advocated parameter growth over data, Chinchilla (2022) equal scaling, later shift to smaller models trained longer (Llama-2-7B 290 tokens/param, Gemma-2-9B 889 tokens/param). Model capability depends on base size, pre-training data, forward computation, and post-training, with post-training still having great potential.