GLM 5.2 local hardware requirements: reported paths by quant

Most "GLM-5.2 local" discussion mixes together different things.

We should separate them before talking about requirements.

1. the full zai-org/GLM-5.2 model
2. low-bit GGUF builds of the full model
3. pruned derivatives like 0xSero/GLM-5.2-504B
4. GGUF builds of that pruned derivative, like 0xSero/GLM-5.2-REAP-504B-GGUF

Those are not the same hardware problem.

At Cyrus, we care about this because coding agents are not short chat demos. They keep repository context around, call tools, recover from mistakes, and spend a lot of time in long prompts. For that workload, "the model loaded" is not the same as "the model is usable."

If someone says "GLM-5.2 runs locally," the next question is: which artifact, which quant, and where are the weights actually sitting?

These are the GLM-5.2-family reports we have been able to tie to specific hardware. They are not all the same benchmark.

	Hardware	Model	Quantization	Technical identifier	Size / memory clue	Reported feel	Source
	DGX Station	GLM-5.2-REAP 504B	unknown	GLM-5.2-REAP-504B-GGUF; exact quant not confirmed	if Q4_K_XL, card lists about 325-331GB; if smaller, conclusion changes	about 60 tok/s; no prefill, context, or concurrency yet	0xSero on X
	4x DGX Spark / GB10	full GLM-5.2	4-bit	UD-IQ4_XS	about 365GB across nodes	6.28 tok/s decode at C=1	NVIDIA forum recipe
	4x GB10	full GLM-5.2 with routed-expert prune	4-bit	AWQ INT4 + pruning + MTP	custom optimized recipe	about 20-22 tok/s decode	NVIDIA forum recipe
	4x DGX Spark	full GLM-5.2, not REAP	4-bit	NVFP4 hybrid	patched vLLM, tight memory assumptions	14.5-15.2 tok/s decode, 450-512 tok/s prefill	NVIDIA forum recipe
	one or two M3 Ultra 512GB systems	full GLM-5.2	4-bit	NVFP4	512GB unified memory system(s)	18.8 tok/s on one, 23.4 tok/s on two; basic decode test	Ivan Fioravanti on X
	2x RTX PRO 6000 Blackwell plus 1TB DDR5	full GLM-5.2	4-bit	UD-Q4_K_XL	1TB DDR5 in the system; not all weights live in VRAM	13-15 tok/s decode, 64 tok/s prefill	Samuel Cardillo on X
	4x RTX 3090 plus 192GB DDR5	full GLM-5.2	2-bit	UD-IQ2_M	223GB on disk; 96GB VRAM plus host RAM	about 7.3 tok/s decode, about 135 tok/s prefill	Reddit report
	2x RTX 5090 plus 512GB DDR5 ECC	full GLM-5.2	5-bit	UD-Q5_K_S	492GB weights; high-RAM workstation	about 12 tok/s	Reddit report
	Dell PowerEdge R740, dual Xeon, 768GB RAM	full GLM-5.2	2-bit	UD-Q2-K_XL	CPU-only, 768GB RAM	4-5.5 tok/s basic chat, about 3 tok/s in opencode	Reddit report

The pattern is clear. Low-bit artifacts can run on varied hardware. Higher-quality artifacts require either a lot of GPU memory, a lot of unified memory, or a willingness to put host RAM in the hot path.

The practical questions are: what can we run, how does it feel, and what does it cost to get there?

This is not a shopping guide. Prices move, used hardware is messy, and several of these runs depend on custom recipes. But the cost shape matters because "GLM-5.2 locally" can mean anything from a slow CPU experiment to a six-figure DGX Station.

That table is why the DGX Station question is not just "is it expensive?" It is whether the machine can turn a much larger coherent-memory box into a better local-agent experience than the cheaper paths above. We split the DGX Station memory profile into its own post because that question depends on the 252GB HBM tier, not just the 748GB headline.

This is the closest thing to a GLM-5.2 requirements table, but it still needs caveats.

The important part is not that the table has big numbers. The important part is that "GLM-5.2 local requirements" changes depending on which row you mean.

We also have two adoption and benchmark context points that are worth keeping separate from hardware requirements: 0xSero said the REAP work reached 15,000 downloads in 10 days and that Zai highlighted it at AI Engineer, and the REAP 504B model card reports 70.5% on Terminal-Bench 2.1 full-89. Those are not throughput numbers.

Sentdex has a practical segment starting around 23:16, and we are treating it as one case study rather than the anchor for the article. The parts worth pulling forward are narrow: he walks through the 8-bit memory math around 26:23, says Q4_K_XL is the tier he would want if it fits around 27:24, and says the Q3_K_XL tier he tried did not feel like the same class of model around 27:57.

There are two separate requirements.

First: enough memory to load the weights.

Second: enough fast memory and runtime headroom to serve the workload.

The second requirement is the one that gets lost. A 245GB 2-bit full-model GGUF and a 325GB Q4_K_XL REAP GGUF are both "local GLM-5.2" to a search engine. They are not the same thing to a machine.

For coding agents, we will measure at least:

We will bring more of these numbers as we collect them. The hard part is not running one prompt. The hard part is getting comparable runs with the exact model artifact, runtime, quant, context, cache settings, and hardware disclosed.

If you want the full model name, start with a full-model GGUF from the Unsloth path and choose the quant tier that fits your memory budget. The public memory table says 223GB for 1-bit, 245GB for 2-bit, 372-475GB for 4-bit, and 570GB for 5-bit.

If you want the 0xSero REAP 504B GGUF path, the Hugging Face model card gives two llama.cpp shapes.

If you install through the llama.app installer shown on the model card, the command uses the unified llama launcher:

llama serve -hf 0xSero/GLM-5.2-REAP-504B-GGUF:Q4_K_XL

or:

llama cli -hf 0xSero/GLM-5.2-REAP-504B-GGUF:Q4_K_XL

If you install llama.cpp from a prebuilt binary, source build, or Homebrew, the executable names are usually explicit:

llama-server -hf 0xSero/GLM-5.2-REAP-504B-GGUF:Q4_K_XL

or:

llama-cli -hf 0xSero/GLM-5.2-REAP-504B-GGUF:Q4_K_XL

We verified locally that llama-cli and llama-server both accept -hf as the Hugging Face repo argument. The command does not make Q4_K_XL fit. It only names the artifact. The listed size is about 325GB before KV cache and runtime overhead.

We would not claim that REAP 504B equals full GLM-5.2. The 70.5% Terminal-Bench 2.1 full-89 number is lower than the full-model numbers we have seen, and the harnesses are not identical.

We would not claim that a 2-bit run tells you what a 4-bit run will feel like.

We would not claim that total memory is enough. That is why the DGX Station memory question matters: DGX Station has 748GB coherent memory, but only 252GB is HBM3e.