GLM 5.2 FP8 — Hardware Requirements & GPU Compatibility
ChatFunctions
GLM 5.2 FP8 is a 753.3B-parameter open language model from Z.ai in the GLM 5 family. It supports a context window of up to 1,048,576 tokens. At Q4_K_M it needs about 456.22 GB of VRAM — see which GPUs and Macs can run it below.
GLM 5.2 FP8 runs on a 512 GB+ Mac or a multi-GPU server
Best local pick → Mac Studio (M3 Ultra, 512GB) · Q4_K_M · 453.8 GB · ~17 tok/s
Macs from 512 GB (Q4_K_M · ~17 tok/s)
MIT
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run GLM 5.2 FP8
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve zai-org/GLM-5.2-FP8 --max-model-len 16384 --tensor-parallel-size 8Size the nodes from the table above
Download: Weights on Hugging Face
Other ways
each opens step-by-stepRun in cloud
Rent 7× H100 SXM · Q4_K_M
- Estimated hourly price
- ~$18.83/hr
- est., whole server
- GPUs
- 7 GPUs
- 560 GB of VRAM
How Much VRAM Does GLM 5.2 FP8 Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 324.4 GB | 323.4 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Runs on 5× H100 SXM 80 GB | |
| 371.5 GB | 370.5 GB | 3-bit medium — 3-bit medium quantization | Runs on 6× H100 SXM 80 GB | |
| 456.2 GB | 455.2 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Runs on 7× H100 SXM 80 GB | |
| 541.0 GB | 540 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Runs on 8× H100 SXM 80 GB | |
| 625.7 GB | 624.7 GB | 6-bit, very good — 6-bit quantization, very good quality | Runs on 5× H200 SXM 141 GB | |
| 757.6 GB | 756.6 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Runs on 6× H200 SXM 141 GB | |
| 1510.9 GB | 1509.9 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Runs on 3 nodes of 8× H100 SXM |
Selected Q4_K_M
4-bit medium quantization453.8GBat 16K context
- Weights
- 452.0 GB
- KV cache
- 1.5 GB
- Overhead
- 0.3 GB
Server-class: ~453.8 GB at 16K context · about 7× H100 SXM 80 GB
Which Hardware Can Run GLM 5.2 FP8?
Q4_K_M · 456.2 GBGLM 5.2 FP8 (Q4_K_M) requires 456.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 594+ GB is recommended. Using the full 1049K context window can add up to 90.3 GB, bringing total usage to 546.5 GB. No consumer GPU has enough memory. 1 device with unified memory can run GLM 5.2 FP8, including Mac Studio (M3 Ultra, 512GB).
Needs server-class hardware
GLM 5.2 FP8 runs on a 512 GB+ Mac or a multi-GPU server · Q4_K_M needs ~453.8 GB
| Hardware | Quant | Memory | Speed | Context | Hardware page | |
|---|---|---|---|---|---|---|
| Mac 512 GB · Mac Studio (M3 Ultra, 512GB) | Q4_K_M | 453.8 GB | ~17 tok/s | 512K ctx | Mac Studio → | Q4_K_M · 453.8 GB · ~17 tok/s · 512K ctx |
| 7× H100 SXM 80 GB · 560 GB | Q4_K_M | 453.8 GB | — | 512K ctx | H100 SXM → | Q4_K_M · 453.8 GB · 512K ctx |
| 4× H200 SXM 141 GB · 564 GB | Q4_K_M | 453.8 GB | — | 512K ctx | H200 SXM → | Q4_K_M · 453.8 GB · 512K ctx |
| 3× B200 192 GB · 576 GB | Q4_K_M | 453.8 GB | — | 512K ctx | B200 → | Q4_K_M · 453.8 GB · 512K ctx |
Also: Mac Studio (M3 Ultra, 512GB) · NVIDIA DGX A100 640GB · NVIDIA DGX H100
By VRAM tier · typical GPUs and Macs at Q4_K_M
Consumer GPUs (up to 32 GB) and Macs up to 256 GB: too small for this model (needs ≥ 321.9 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
Alternatives to GLM 5.2 FP8
GLM 5 family · 6 models →Similar models
3| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| DeepSeek V4.1 FlashDeepSeek | 129.5 | 458.4 GB | —needs > 24 GB | same size |
| Atria Dawn PreviewInternLM | — | 1508.4 GB | —needs > 24 GB | same size |
| Hy4 PreviewTencent | — | 1564.2 GB | —needs > 24 GB | same size |
Fine-tunes & variants(2)
Specifications
- Publisher
- Z.ai
- Family
- GLM 5
- Parameters
- 753.3B
- Active parameters
- 41.8B per token
- Architecture
- GlmMoeDsaForCausalLM
- Context Length
- 1,048,576 tokens
- Vocabulary Size
- 154,880
- Release Date
- 2026-06-16
- License
- MIT
Frequently Asked Questions
How much VRAM does GLM 5.2 FP8 need?
GLM 5.2 FP8 requires 456.2 GB of VRAM at Q4_K_M, or 1510.9 GB at BF16. Full 1049K context adds up to 90.3 GB (546.5 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 753.3B × 4.8 bits ÷ 8 = 452 GB
KV Cache + Overhead ≈ 4.2 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 94.5 GB (at full 1049K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run GLM 5.2 FP8?
No — GLM 5.2 FP8 requires at least 324.4 GB at Q2_K, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
What's the best quantization for GLM 5.2 FP8?
For GLM 5.2 FP8, Q4_K_M (456.2 GB) offers the best balance of quality and VRAM usage. Q5_K_M (541.0 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 324.4 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run GLM 5.2 FP8 on a Mac?
Yes — Mac Studio (M3 Ultra, 512GB) and above can run GLM 5.2 FP8. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 456.2 GB of usable unified memory (RAM minus macOS overhead).
Can I run GLM 5.2 FP8 locally?
Yes — GLM 5.2 FP8 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
What's the download size of GLM 5.2 FP8?
At Q4_K_M, the download is about 452.00 GB. The full-precision BF16 version is 1506.66 GB. The smallest option (Q2_K) is 320.17 GB.
Which GPUs can run GLM 5.2 FP8?
No single consumer GPU has enough VRAM to run GLM 5.2 FP8 at Q4_K_M (456.2 GB). Multi-GPU or professional hardware is required.
Which devices can run GLM 5.2 FP8?
1 device with unified memory can run GLM 5.2 FP8 at Q4_K_M (456.2 GB), including Mac Studio (M3 Ultra, 512GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.