What's this about?
OrangeAI's observation about Zhipu's GLM 5.3 is more interesting than the release itself: the base model is identical to GLM 5.2's 756B parameters. The entire capability jump came from post-training — and Zhipu says considerable headroom remains. Read alongside DeepSeek V4 Flash beating GLM 5.2 with only 304B parameters, his conclusion is that post-training is where the gains now live, and he suspects OpenAI's Luna is only around 500B parameters.
What exactly is "post-training"?
Pre-training is the expensive part: feeding a model enormous quantities of text until it learns language and world knowledge. That produces something knowledgeable but not useful — it completes text rather than helping you. Post-training is everything after: supervised fine-tuning on good examples, reinforcement learning from human or automated feedback, and training on reasoning traces. It's a fraction of the compute of pre-training. The significance of the same base model improving substantially is that capability gains no longer require the enormous expense of training from scratch.
Why is this worth noticing?
- If post-training carries the gains, the barrier to competing drops sharply — you need one good base model and then iterate cheaply, which is roughly the Chinese labs' current strategy.
- It undermines parameter count as a proxy for capability. A 304B model beating a 756B one, and the same 756B improving substantially without changing, both point the same way.
- The Luna speculation is inference, not fact, and should be read that way — but the reasoning behind it (frontier results don't require frontier size if post-training is strong) is the substantive claim.