Z.ai released GLM-5.3 this week. Parameter scale, architecture, and active parameters (the parameters that actually participate in computation on each inference — MoE models distinguish these two) are identical to its predecessor GLM-5.2. The only change: one extra month of long-chain RL (reinforcement learning — letting AI practice repeatedly in simulated environments and self-adjust via reward signals) training. We're not interested in the release itself, but in the claim behind it: once parameters hit a threshold, capability is decided by post-training, not by model size.

What this is

Z.ai (Zhipu)'s experiment is a textbook controlled variable: the model body stays fixed, and the only variable is post-training time and method. On certain long-chain tasks (problems requiring multi-step reasoning that can't be solved in one shot — like finding code vulnerabilities), GLM-5.3 is clearly stronger than GLM-5.2.

The piece places this result in a larger historical arc:

  • In 2020, Kaplan's paper argued parameters should scale 2.7x faster than data. The industry followed suit and produced a wave of "big-but-dumb" models;
  • In 2022, Hoffmann redid the experiment with 400 models and concluded each parameter paired with 20 tokens is most cost-effective — the Chinchilla law;
  • Once models were called billions of times per day, inference cost (compute spend per query) overtook training cost as the norm, and the "small model, fed with data" path of Llama-2-7B and Gemma-2-9B took off;
  • With the MoE (Mixture of Experts — splitting one large model into multiple smaller modules, activated on demand) architecture, the picture shifts again — total parameters determine how much the model can "memorize," active parameters determine how many reasoning steps it can take, and the two can't be conflated.

Z.ai cites Roberts et al.'s 2025 research pointing out a key distinction: rote-memorization tasks get better with more parameters, while reasoning tasks get better with more data. The equation "parameters = capability" never held.

Industry view

Supporters call it "a sober step." A researcher long focused on model evaluation told the LocalLLaMA subreddit that over the past two years, the industry poured 90% of its money into parameter growth, yet 90% of user pain points with AI are "it can't think through problems" — not "it doesn't know things." Z.ai's controlled experiment proves post-training has independent value — the first time the industry has put that evidence on the table in public.

Opposition is equally sharp. An investor who has long tracked Chinese large-model companies told us the blog reads more like PR than science — it discloses only "what was done," not "which baseline was used as control" (baseline — the comparison group required for any experiment), "whether the eval set is independent" (benchmark — a public test set), or "how much cost went up." A month of RL training's real compute bill may exceed that of swapping in a bigger base model. A good model doesn't mean enterprises will pay for it.

There's another hidden concern: long-chain RL training depends on high-quality simulated environments. Fewer than ten companies globally can build such environments. If post-training becomes the new moat, compute advantages will concentrate more than in the parameter era — players outside OpenAI, Anthropic, Google, Z.ai, and DeepSeek may not even get an entry ticket. The "democratization" narrative around large models is quietly reversing.

Impact on regular people

  • For enterprise IT: Over the next 18 months, when procuring large-model APIs (call interfaces, billed by token), "parameter scale" will no longer be the primary metric. Attention will shift to "task benchmark scores" and "per-unit inference quality." Existing procurement templates need rewriting.
  • For individual careers: Practitioners who can write prompts and decompose task chains continue to rise in value; the space for building advantage purely on "I understand parameters" is shrinking.
  • For the consumer market: Next year, consumer AI products will see a wave of experience stratification — at the same price point, some will be "knowledgeable but answering the wrong question," others "think deep but respond slowly." Buy by task type, not by model name.