What Is GLM 5.3? The Coding-First Release That Skips a New Base Model

Quick answer: GLM 5.3 is Z.AI's new flagship release, announced August 14, 2026. It runs on the exact same base model as GLM 5.2 — every improvement comes from scaled post-training, not new pretraining. The result: the strongest open-weights coding model yet (a 50% jump on Z.AI's internal code bench), open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam, and surprising emergent cybersecurity capability. Weights land on HuggingFace two weeks after launch.

TL;DR

What Is GLM 5.3?

GLM 5.3 is the newest model in Z.AI's GLM family, released on August 14, 2026. The headline detail that makes it unusual: it is not a new base model. Z.AI kept the GLM 5.2 base and spent the last month scaling post-training on the stack it built for 5.2 — IndexShare for long-context processing, SAO for RL on long-horizon tasks, and slime for large-scale asynchronous training.

"Scaling post-training is all we did for GLM-5.3," Z.AI writes. Everything users notice — sharper coding, longer agent runs, more reliable tool use — comes from more environments, more diverse tasks, and more compute spent on reinforcement learning, not from a bigger pretraining run.

That means three practical things for developers:

1. API compatibility is nearly drop-in — the same endpoint, same context behavior, with a small thinking-parameter change (below).
2. The gains are concentrated where RL can help — coding, agents, tool use, long-horizon work. Not a general knowledge dump.
3. It stays open — weights are coming to HuggingFace two weeks after launch, after safety evaluation and hardening.

GLM 5.3 vs GLM 5.2: What Actually Changed

Since the base is identical, the honest way to read GLM 5.3 is as "GLM 5.2, but trained harder." The official benchmark table shows where the extra training landed:

Coding benchmarks:

The Terminal-Bench 3.0 jump (4.6 → 28.3) is the standout: it's a genuinely harder benchmark, and GLM 5.3 now sits close to closed frontier models on it. On Z.AI's private Z.ai Code Bench, GLM 5.3 at Max effort hits 34.5% with ~75K output tokens per task, versus GLM 5.2's 23.4% at ~96K — better results with fewer tokens. At High effort it beats Claude Opus 4.8 (31.4% at ~50K tokens vs 29.5% at 120K).

The unexpected part — cyber:

CyberGym is the best public result on the benchmark, ahead of Mythos 5 and GPT-5.6 Sol. Z.AI is transparent that this capability emerged faster than expected while scaling post-training — and that it's a dual-use reason they're holding weights for two weeks of safety evaluation.

How to Use GLM 5.3 Today

Three ways, in order of convenience:

1. GLM Coding Plan — Z.AI rolled GLM 5.3 out to all Coding Plan users on launch day. The plan now uses a points-based quota: input, cached input, and output tokens are metered separately, and calls outside peak hours (14:00–18:00 UTC+8, Mon–Fri) cost 50% of standard points.
2. ZCode — Z.AI's coding agent with a 98%+ cache hit rate, a limited-time 1.5× quota boost (through August 31), Goal mode for long-horizon tasks, and Remote Control to steer jobs from WeChat or Feishu.
3. API — model id glm-5.3 on the Z.AI platform, with the thinking parameters below.

The One API Change You Must Know

GLM 5.3 no longer supports disabling thinking. thinking.type: "disabled" now fails the request. Migration:

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

reasoning_effort accepts low, high, or max (default max). Z.AI recommends max for coding tasks. If you previously sent thinking.type: "disabled", switch to enabled with reasoning_effort: "low" before bumping the model ID, or the request errors.

Why It Matters

GLM 5.3 matters for three reasons beyond the scoreboard:

• It proves the post-training playbook. Same base, meaningfully better model — a signal that RL scaling on long-horizon tasks is still an unlocked lever, not a solved one.
• It raises the open-weights ceiling again. Two weeks out from weights, GLM 5.3 will be the strongest open model for coding and agentic work, self-hostable on your own hardware.
• It forces the safety conversation. When a model's exploitation capability more than doubles in one release, the "open weights" debate stops being theoretical — which is exactly why Z.AI is holding the release until hardening is done.

FAQ

Is GLM 5.3 a new base model? No. It shares the GLM 5.2 base model; all gains come from scaled post-training (more RL environments, more compute).

When will GLM 5.3 weights be released? Two weeks after the August 14 launch, once safety evaluation and hardening complete — "coming soon" on HuggingFace.

Is GLM 5.3 free? GLM Coding Plan subscribers get it immediately; off-peak usage consumes 50% of standard points. The Z.AI API bills per token.

Do I need to change my API code? Only if you used thinking.type: "disabled" — that's no longer supported. Set enabled and pick reasoning_effort: low, high, or max.

Is GLM 5.3 multimodal? Like GLM 5.2, it's text-in, text-out — no vision or audio input.

Is GLM 5.3 good for coding? It's the strongest open-weights coding model on Z.AI's own benchmarks, with open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam.

Sources

• Z.AI: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities (August 14, 2026)

Last updated: August 14, 2026