GLM 5.3 Architecture: Same Base as 5.2, ~743B MoE — Post-Training Explained

Quick answer: GLM 5.3 is not a new pretraining run — it uses the same base model as GLM 5.2 (a ~743–753B-parameter Mixture-of-Experts model by community estimates; Z.AI hasn't published an exact count). Every improvement comes from a second phase: massive post-training — more long-horizon RL environments, more task diversity, and more compute on Z.AI's IndexShare + SAO + slime stack. Same weights base, dramatically better behavior.

TL;DR

Spec	GLM 5.3
Base model	Same as GLM 5.2
Estimated size	~743–753B MoE (community estimate — not official)
Context window	1M tokens
Max output	128K tokens
Modality	Text only
Training	Post-training only (no new pretraining)
Key stack	IndexShare, SAO, slime
Open weights	Yes — in ~2 weeks

The One-Line Architecture Story

"Scaling post-training is all we did for GLM-5.3."

That's Z.AI's own summary, and it's the most important architectural fact about the model: the pretrained base is identical to GLM 5.2's. The 743–753B MoE figure floating around the community is an estimate of that shared base — Z.AI has not published an exact parameter count, and until the weights drop, treat any specific number as unconfirmed.

What "Same Base" Means Technically

If you ran GLM 5.2 in production, GLM 5.3 inherits its serving profile:

1M-token context window, 128K max output — identical to 5.2.
Text in / text out — not multimodal (same as 5.2).
Same MoE architecture — sparse activation, so serving cost and quantization strategy carry over.
Same function-calling and tool-use surface — with one API change (thinking can no longer be disabled).

The practical consequence: if you self-host GLM 5.2, your serving stack (SGLang/vLLM, sharding, quantization) serves GLM 5.3 with minimal changes.

The Post-Training Stack That Made the Gains

Z.AI credits three pieces of infrastructure for GLM 5.3's improvements — all built during the GLM 5.2 cycle and scaled over the past month:

Component	Role
IndexShare	Efficient long-context processing (the 1M-token window that makes repo-scale work possible)
SAO	RL on long-horizon tasks — the training method behind agentic gains
slime	Large-scale asynchronous training framework — open-source, powers the rollout

The training loop: more environments, more diverse tasks, more compute on this stack. Z.AI reports the system-level optimizations improved end-to-end RL training throughput 2.3×+ for long-horizon coding tasks — that's how a one-month post-training sprint produces Terminal-Bench 3.0 gains (4.6 → 28.3).

How Post-Training Changed the Behavior

Same base, different model:

Coding: +50% on Z.AI's internal code bench; open-weights SOTA on Terminal-Bench 3.0 and Agents' Last Exam.
Agents: AutomationBench nearly doubled (26.2 → 48.2).
Cyber (emergent): CyberGym 84.5 (best public result) — the capability Z.AI says "developed faster than we expected" as training scaled.
Efficiency: better results with fewer output tokens (34.5% at ~75K tokens vs 5.2's 23.4% at ~96K at Max effort).

The One API Change to Remember

Architecture carries over; the thinking interface changed:

{
  "model": "glm-5.3",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "max"
}

thinking.type: "disabled" is no longer supported — GLM 5.3 always reasons.
reasoning_effort: low | high | max (default max).
Migration: switch to enabled + low before updating the model ID.

What This Means for You

If you're on GLM 5.2: upgrade is low-risk — same serving footprint, one parameter change, strictly better benchmarks.
If you're self-hosting: wait for the weights (~2 weeks), keep your 5.2 serving setup, swap the model.
If you're choosing between models: GLM 5.3 proves post-training scaling is a real lever — the "same base, better model" pattern means rapid iteration without new pretraining costs.

FAQ

Is GLM 5.3 a new architecture? No — same base model as GLM 5.2. All gains come from post-training (RL on long-horizon environments).

How many parameters does GLM 5.3 have? Community estimates put the shared base at ~743–753B MoE, but Z.AI hasn't published an exact count. Treat specific numbers as estimates until the weights release.

Does GLM 5.3 have a bigger context window than 5.2? No — both have 1M-token context, 128K max output.

Is GLM 5.3 multimodal? No — text in, text out, like GLM 5.2.

What changed in the API? Thinking is always on: reasoning_effort (low/high/max) replaces the ability to disable thinking.