Z.ai CEO Jie Tang says GLM 5.3 is the same base model as 5.2 plus one month of long-horizon RL, and calls parameter count meaningless on its own
In a Latent Space AINews piece published August 20, Jie Tang lays out five scaling knobs and argues "parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions." His claim about GLM 5.3 is the concrete part: no architecture change from 5.2, roughly a month of extra reinforcement learning on multi-step environments modeled on tasks that would take an experienced engineer several days, including infra diagnostics with live access to clusters, docs and codebases. Tang's point for builders is that capabilities like vulnerability discovery need 20+ step reasoning chains, not more memorized weights, which is why a mid-size open model keeps closing on frontier scores.
Source
↳ Follow the thread