How to Deploy GLM 5.3 Locally
GLM 5.3 is a MoE model with 744B total parameters and 40B active per token. At Q4 it needs 445GB of VRAM to sit entirely on the GPU; with experts offloaded to system RAM it needs only 41GB of VRAM plus 406GB of RAM. The cheapest device that holds it is the Mac Studio M3 Ultra 512GB, with no price data yet, at roughly 11.5 tokens/s. In offload mode the cheapest card that runs it is the Ryzen AI Max+ 395 128GB, at roughly 2 tokens/s.
GLM 5.3 details
How much VRAM does GLM 5.3 need for local deployment?
MoE: all 744B parameters must stay resident in memory, while speed is set by the 40B active per token.
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| 1-bit | 228.5GGUF | 204.9 | 292.6 | 643.6 |
| 2-bit | 253.9GGUF | 306.7 | 394.5 | 745.5 |
| 3-bit | 343GGUF | 386.7 | 474.5 | 825.5 |
| 4-bit | 467.3GGUF | 445 | 532.7 | 883.7 |
| 5-bit | 562.5GGUF | 532.3 | 620 | 971 |
| 8-bit | 801.4GGUF | 801.5 | 889.2 | 1,240.2 |
| FP16 | 1,508GGUF | 1,485.3 | 1,573.1 | 1,924.1 |
Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 1M limit show a dash. Each tier takes whichever file of that bit width the repository has, preferring unsloth dynamic quants, so the exact file name varies by model.
Which GPUs can run GLM 5.3?
AmazonJP
| Device | VRAMGB | Price | Est. t/s | Suggested context |
|---|---|---|---|---|
| Mac Studio M3 Ultra 512GB | 512 | —No listings | 11.5 | 8K |
Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.
Recommended GPUs for GLM 5.3 with CPU/GPU offloading
AmazonJP
Expert weights live in system RAM (needs at least 406 GB); VRAM holds only attention layers, shared experts and the KV cache, at least 41 GB at Q4. Speed is bound by RAM bandwidth (estimated at 70 GB/s) and is far slower than a full-VRAM setup.
| Device | VRAMGB | Price | RAM neededGB | Est. t/s |
|---|---|---|---|---|
 Ryzen AI Max+ 395 128GB | 128 | $3,913 | 406 | 2.2 |
 DGX Spark 128GB | 128 | $8,223 | 406 | 2.5 |
| GeForce RTX 4090 48GB | 48 | — | 406 | 3.3 |
| Radeon PRO W7900 | 48 | — | 406 | 3.1 |
| RTX PRO 5000 Blackwell 48GB | 48 | — | 406 | 3.4 |
| CMP 170HX 64GB | 64 | — | 406 | 3.4 |
| Instinct MI210 | 64 | — | 406 | 3.4 |
| RTX PRO 5000 Blackwell 72GB | 72 | — | 406 | 3.4 |
| RTX PRO 6000D | 84 | — | 406 | 3.4 |
| RTX PRO 6000 Blackwell | 96 | — | 406 | 3.5 |
| Mac Studio M4 Max 128GB | 128 | — | 406 | 2.9 |
| Mac Studio M5 Max 128GB | 128 | — | 406 | 2.9 |
| MacBook Pro M4 Max 128GB | 128 | — | 406 | 2.9 |
| MacBook Pro M5 Max 128GB | 128 | — | 406 | 2.9 |
Deploy GLM 5.3 locally with Ollama, llama.cpp or vLLM
The Ollama library has no official tag for it yet.
llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.
llama-server -hf unsloth/GLM-5.3-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve zai-org/GLM-5.3
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does GLM 5.3 need?
At Q4_K_M with an 8K context GLM 5.3 needs about 445GB of VRAM; 536GB leaves comfortable headroom.
What is the cheapest GPU that runs GLM 5.3?
The Mac Studio M3 Ultra 512GB: 512GB of VRAM at roughly 11.5 tokens/s. We have no price data for it yet.
Can you run GLM 5.3 with Ollama?
Not yet: the Ollama library has no official tag for it. You can load the GGUF with llama.cpp instead.