GLM-5.3 hardware requirements

GLM-5.3 is the strongest open-weight model you can download today. Z.ai launched it on 14 August 2026 and published the weights two weeks later, on 28 August, once safety hardening was finished.⁴ It reports state-of-the-art results on Terminal-Bench 3.0 and lifts that score from 4.6% to 28.3% against its own predecessor.⁴

It is also a 744B-parameter model whose usable build wants 475GB of memory. That number, not the benchmark, decides whether you can run it.

• Parameters744Babout 40B active
• Context1Mtokens
• LicenceGLM-5.3custom, not MIT
• Memory at 4-bit475GB94.3% accuracy

The same model as 5.2, taught further

GLM-5.3 shares its base model with GLM-5.2. Every gain comes from additional post-training rather than a fresh pre-training run.¹ The architecture, the parameter count and the memory bill are therefore identical to 5.2's, which is convenient: anything you learned about running 5.2 still applies, and any machine that ran 5.2 runs 5.3.

Like every frontier open model this year it is a mixture of experts. It holds 744B parameters and routes about 40B of them per token. The small active figure is why a single machine can generate at a reasonable speed. The large total figure is why that machine needs to be a serious one — the router picks different experts for every token, so all of them stay resident.

What it needs

Unsloth publish a full quantisation ladder for GLM-5.3, and unusually they publish a measured accuracy figure beside each one. That second column is what makes the choice possible.

What each quantisation gives up

Quantisation stores each weight in fewer bits. The saving is real and so is the cost, and the cost is not linear. Read the ladder from the bottom and the first few steps are nearly free; read it from the top and the last few steps are brutal.

Going from 4-bit to 2-bit saves 230GB and costs nearly sixteen points of accuracy. Going from 5-bit to 4-bit saves 95GB and costs one and a half. If your budget stretches to 512GB, 4-bit is where the model still behaves like itself. Below that you are running a demonstration.

Three ways to reach 475GB

One pool of unified memory. A Mac Studio M3 Ultra 512GB holds the 4-bit build with headroom for context, and it is the machine that does this today. The 2-bit build at 245GB fits a 256GB machine directly.² 512GB is the ceiling on Apple silicon either way: the newer M5 Ultra tops out at the same figure, and its 512GB configuration is not due until late October.⁵

A card plus system RAM. Because the model is sparse, the split works well. Attention layers go on the graphics card, expert layers stay in system memory, and llama.cpp's --n-cpu-moe flag sets where the line falls. Unsloth report GLM-5.2 running on a single 24GB card with 256GB of RAM this way.² Expect it to be slower than unified memory, because the expert weights then read at the speed of your DDR5 rather than your VRAM. This is where an eight-channel workstation platform earns its price.

Everything in VRAM. Four RTX PRO 6000 cards give 384GB, enough for a 3-bit build. HelixML measured the smaller GLM-5.3-Flash on four RTX PRO 6000 Server Edition cards at 63.7 output tokens per second for a single request and 82.8 with speculative decoding, and needed a pinned experimental build of SGLang to do it.³ Day-zero support for a new architecture is rarely finished.

Read the licence before you plan a product on it

This is where GLM-5.3 differs from its own family, and almost nobody mentions it. GLM-5.2 was released under MIT.⁶ GLM-5.3 is not: its weights carry a bespoke licence named after the model, which requires model-as-a-service operators above $10B in trailing twelve-month revenue to pass a Z.ai security review before commercial use.⁴ GLM-5.3-Flash, released on 26 August, is MIT again.⁴

For most businesses the revenue threshold is academic. The point is that the licence now varies release by release inside one family, so check it per model rather than per vendor.

Settings that matter

Use a temperature of 1.0 with top-p at 0.95 for ordinary work, and top-p at 1.0 for long agentic runs.¹

One trap is specific to this release. GLM-5.3 uses maximum reasoning by default and thinking cannot be switched off, which is not true of 5.2. Pass --chat-template-kwargs '{"reasoning_effort":"low"}' to bring it down, and set clear_thinking to true for multi-turn conversations so old reasoning does not accumulate in the context.¹ Left alone on a long conversation it will spend both your tokens and your memory headroom.

For the arithmetic behind all of this, see how much memory you need to run an open model locally. For the newest release in the same weight class, see DeepSeek V4.1-Flash. A 512GB machine today means a Mac Studio M3 Ultra 512GB or a Neox Metal with 512GB of registered ECC memory.