ModelsGLM-5.3 Guide: A Post-Training Release With Breaking Changes
GLM-5.3 keeps the GLM-5.2 base and gains everything from post-training. The benchmark jump, the always-on reasoning change, and the licence to read before shipping.
GLM-5.3 is the rare release where the parameter count tells you nothing new. Z.ai shipped it in mid-August 2026 on the same base model as GLM-5.2, and its documentation is explicit that every gain comes from post-training. The model card lists 753B parameters, the same class as its predecessor.
That makes it a clean case study in how far training alone can move a model. It also makes it easy to miss the two changes that will actually affect your integration: reasoning can no longer be disabled, and the weights no longer ship under MIT.
Same base, different post-training
Because the base is unchanged, the serving profile is too. Context is 1M tokens, maximum output is 128K, and input is text only. Hardware that served GLM-5.2 serves GLM-5.3; the model card lists vLLM, SGLang, Transformers, KTransformers and Unsloth support.
What moved is behaviour on long agentic work. Z.ai's model card compares the two directly:
- Terminal-Bench 2.1 — 88.2, up from 81.0
- Terminal-Bench 3.0 — 28.3, up from 4.6
- DeepSWE v1.1 — 66.9, up from 46.2
- Toolathlon Verified — 73.0, up from 59.9
- HLE with tools — 62.5, up from 54.7
- CyberGym — 84.5, up from 77.2
- GDPval-AA v2 — 1,769, up from 1,508
Z.ai also reports roughly a 50% improvement over GLM-5.2 on its internal Z.ai Code Bench, and describes the model as state-of-the-art among open models on Terminal-Bench 3.0. The same comparison table shows it behind the top closed model on Terminal-Bench 3.0 and DeepSWE, so this is an open-weight lead, not a frontier lead. As always, these are vendor figures on vendor scaffolds, and a different harness can move a Terminal-Bench score by several points.
Reasoning can no longer be switched off
This is the change that breaks code. On GLM-5.3, reasoning is always on. The thinking.type parameter only accepts enabled, and Z.ai's migration note says requests that send disabled will fail. The replacement is reasoning_effort, with three levels: low, high and max. The default is max.
Two consequences follow. First, any harness that disabled thinking for cheap turns needs changing before you switch model ids, or those calls error. Second, the default is the most expensive setting. If you migrate without touching effort, every call reasons at maximum depth.
Z.ai's own figures make the trade concrete. On Z.ai Code Bench, GLM-5.3 scores 34.5% at max effort using about 75K output tokens per task, and 31.4% at high effort using about 50K. A third fewer output tokens for three points is a trade most production workloads should take by default and revisit only where measured. Hidden reasoning output is where the bill lives on a model like this.
Check the licence before you ship
GLM-5.1, GLM-5.2 and the smaller GLM-5.3 Flash are MIT-licensed on Hugging Face. The GLM-5.3 model card lists its licence as glm-5.3 — a named custom licence, not MIT.
The model card does not reproduce the terms, so this guide will not characterise them. The practical point is that the assumption carried over from earlier GLM releases — fine-tune, redistribute and resell without asking — may no longer hold. If you serve inference to customers or redistribute weights, read the licence file itself before building on it. Open-weight model licences compared lists the clauses that usually matter.
What it costs
Z.ai's pricing page lists GLM-5.3 at $1.40 per million input tokens, $0.26 cached and $4.40 output — unchanged from GLM-5.1 and 5.2. At the time of writing, the OpenRouter listing showed an input rate far below that ($0.079 per million) next to $4.80 output; a split that lopsided is worth confirming with the provider before you budget on it.
OpenRouter also lists a separate :batch endpoint for GLM-5.3, at $0.45 input and $2.00 output per million tokens at the time of writing, with a 128K output cap. Batch routes generally trade latency for rate, which suits evaluation sweeps and offline jobs rather than interactive agents; check the provider's turnaround terms before relying on it.
Choosing within the 5.3 family
GLM-5.3 now has three siblings, and the right one depends on what you are short of.
- Short of money, or need images and video: GLM-5.3 Flash is a smaller, natively multimodal model at a fraction of the price, and it comes within a few points of the full model on several agentic benchmarks.
- Short of time on the full model: GLM-5.3 Prime is listed as a faster-serving variant of this model at a higher price.
- Short of time on the small model: GLM-5.3 FlashX is the high-speed tier of Flash.
Pick full GLM-5.3 when the task is long, text-only and hard enough that the gap to Flash shows up in your own results — security review and multi-hour terminal work are the obvious cases.
Migration checklist
- Search your code for
thinkingset to disabled and replace it withreasoning_effort: "low". - Set an explicit effort level rather than inheriting
max. - Read the
glm-5.3licence if you redistribute or resell. - Rerun your own task set at high and max effort and compare cost per solved task, not pass rate alone.
Common questions
Is GLM-5.3 a new architecture?
No. Z.ai states it uses the same base model as GLM-5.2 and that every improvement comes from post-training. Context stays at 1M tokens and maximum output at 128K, so serving requirements are unchanged.
Can I turn off reasoning on GLM-5.3?
No. Reasoning is always on, and requests that set thinking to disabled will fail. Use reasoning_effort set to low for cheap turns; the levels are low, high and max, and max is the default.
Is GLM-5.3 MIT licensed like GLM-5.2?
Its Hugging Face model card lists a custom licence named glm-5.3 rather than MIT. Read the licence text before redistributing the weights or serving them commercially, rather than assuming earlier GLM terms carry over.
