Z.ai released GLM-5.3, a controlled experiment demonstrating that scaling post-training yields significant gains in long-horizon reasoning without changing the base model or parameter count.
The release keeps the same architecture and total/activated parameters as GLM-5.2 but adds one month of scaling on long-horizon environments and reinforcement learning. The authors argue that effective depth and post-training are now more critical than parameter count for capabilities like carrying complex inference chains. Scaling laws have shifted from optimizing training compute to balancing data, compute, and inference costs, with optimal tokens-per-parameter varying by task.
This approach allows Z.ai to improve specific capabilities without increasing model size, reserving base model scaling for future iterations.