Hardware guide · October 2026
What GPU do you need to run GLM-5.3-Flash?
GLM-5.3-Flash is a 320B (18B active) mixture-of-experts model. Here is exactly how much memory it needs and which GPUs and Macs run it well.
The short answer
To run GLM-5.3-Flash at Q4_K_M with 8K tokens of context you need about 201GB of VRAM (194.0GB of weights plus KV cache and overhead). No single consumer GPU can hold it. The most affordable setup is 8× RTX 5090 (~29 tokens/sec). On a Mac or other unified-memory system, the M3 Ultra 512GB runs it at ~17 tokens/sec.
- VRAM at Q4_K_M
- ~201GB
- GPUs that run it
- 3
- Parameters
- 320B
- CPU only (est.)
- ~4.1 tok/s
Our recommendations
The best GPUs for GLM-5.3-Flash
Multi-GPU
8× RTX 5090256GB total · ~$34,400
- Quant
- Q4_K_M
- Speed (est.)
- ~29 tok/s
Split across cards with llama.cpp, vLLM, or ExLlama. Needs a board and PSU that can feed every card.
Mac / unified memory
M3 Ultra 512GB512GB unified · $9,499 MSRP
- Quant
- Q6_K
- Speed (est.)
- ~17 tok/s
Lots of memory per dollar and silent, but slower than a GPU with the same capacity.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Memory
GLM-5.3-Flash VRAM requirements
Memory needed at each quantization level. Q4_K_M is the best balance of size and quality for most people; move up to Q5 or Q6 if you have the room.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| Q2_K | Rough | 120 GB | 124 GB | 125 GB |
| Q3_K_M | Usable | 156 GB | 161 GB | 162 GB |
| Q4_K_M | RecommendedDefault | 194 GB | 201 GB | 201 GB |
| Q5_K_M | Very good | 228 GB | 236 GB | 236 GB |
| Q6_K | Near-lossless | 264 GB | 273 GB | 273 GB |
| Q8_0 | Practically lossless | 340 GB | 351 GB | 351 GB |
| BF16 | Full precision | 640 GB | 660 GB | 660 GB |
Context
How context length changes the memory you need
Every token in the context window is stored in the KV cache. At Q4_K_M, with a 16-bit KV cache, GLM-5.3-Flash needs:
| Context | KV cache | Total memory | With 8-bit KV |
|---|---|---|---|
| 4K tokens | 0.0 GB | 200.5 GB | 200.4 GB |
| 8K tokens | 0.1 GB | 200.5 GB | 200.5 GB |
| 16K tokens | 0.2 GB | 200.6 GB | 200.5 GB |
| 32K tokens | 0.4 GB | 200.8 GB | 200.6 GB |
| 64K tokens | 0.7 GB | 201.2 GB | 200.8 GB |
| 128K tokens | 1.5 GB | 201.9 GB | 201.2 GB |
| 256K tokens | 3.0 GB | 203.4 GB | 201.9 GB |
Hybrid: 34 of 45 layers use Kimi Delta Attention with a fixed-size state, and only 11 layers use sparse MLA with a growing cache. Assuming a 512-value compressed latent per MLA layer as in GLM-5, with no RoPE part because these layers use NoPE: 11 x 512 x 2 bytes = 11,264 bytes per token, about 11.8GB at 1M tokens. Not yet confirmed against a runtime log.
Compatibility
Every GPU that runs GLM-5.3-Flash
Each GPU at the highest-quality quantization that fits with 8K context, fastest first. Speeds are single-user estimates.
| GPU | Memory | Quant | Speed | Fit | Price |
|---|---|---|---|---|---|
| M5 Ultra 256GB | 256GB unified | Q3_K_M | ~39 tok/s | Runs well | ~$9,499 |
| M5 Ultra 512GB | 512GB unified | Q6_K | ~25 tok/s | Runs well | — |
| M3 Ultra 512GB | 512GB unified | Q6_K | ~17 tok/s | Runs well | $9,499 |
Tips
Getting the most out of GLM-5.3-Flash
- Start at Q4_K_M. It keeps nearly all of the quality at about a third of the 16-bit size. Step up to Q5_K_M or Q6_K only if it still fits with the context you need.
- Budget memory for context. The KV cache grows with every token. Quantizing it to 8-bit roughly halves that cost with little quality loss.
- Offload experts, not layers. For mixture-of-experts models, llama.cpp's
--n-cpu-moe(or-ottensor overrides) keeps attention on the GPU and moves expert weights to system RAM. That keeps speed usable when the model does not fit in VRAM. - Pick the right runtime. Ollama and LM Studio are easiest. llama.cpp gives the most control. vLLM or SGLang are best when several people or agents share one machine.
Want to try different quantizations, context lengths, or GPU counts? Use our VRAM calculator, or read the GLM-5.3-Flash overview for setup commands and benchmarks.
Frequently asked questions
How much VRAM do I need to run GLM-5.3-Flash?
About 201GB at Q4_K_M with an 8K context window. At Q8_0 (near-lossless) it needs about 351GB, and at Q3_K_M about 161GB with some quality loss. Longer contexts add 0.01GB per 1K tokens with a 16-bit KV cache.
Can I run GLM-5.3-Flash on a 24GB GPU?
No. GLM-5.3-Flash needs about 201GB at Q4_K_M, which does not fit on a 24GB card like the RTX 3090 or RTX 4090, even with CPU offload on a 64GB system.
Can I run GLM-5.3-Flash on a Mac?
Yes. With 512GB of unified memory or more (for example the M3 Ultra 512GB), macOS can give the GPU enough memory to run it at Q6_K, at about ~17 tokens/sec. MLX builds are usually fastest on Apple silicon.
Can I run GLM-5.3-Flash on CPU only, without a GPU?
Technically yes on a server with enough RAM, but at about 201GB it needs a multi-channel server platform and will be slow on a desktop.