GLM VRAM requirements: can I run GLM-5.3 or GLM-5.2 locally?
GLM-5.3 and GLM-5.2 are 754 billion parameters each, 467 GB at a 4-bit quant. Here's what that means for a Mac Studio, and the two Flash models that fit far smaller machines.
Check your machine
Won't fit 201 GB
Weights, cache and overhead add up to more than the card and the RAM together.
- Weights 200 GB
- Context cache 0.6 GB
- Overhead 0.8 GB
- Fast memory
- With system RAM
What we found
GLM-5.3 and 5.2 are the same size, so one answer covers both. At 4-bit that's a 512 GB Mac Studio with its GPU memory limit raised, and the smallest file, UD-IQ1_S (217 GB), still comes to 221 GB with a 32K context.
GLM-5.3 Flash is where it gets interesting. Three quarters of its layers use linear attention, so a million tokens of context costs 14.9 GB of cache. Full GLM-5.3 would need 115 GB for the same million. Its smallest file, UD-IQ1_S (93.1 GB), fits a DGX Spark or a 128 GB Mac.
For a single 24 GB card, GLM-4.7 Flash is the GLM that fits: 20.9 GB with a 32K context. Like DeepSeek it caches a compressed latent instead of full keys and values, which is why its cache stays small.
One detail for the GLM-5 models: their sparse attention keeps an extra indexer key for every token, and we count it in the cache.
GLM-5.3 Flash
321 billion parameters · mixture of experts, 8 of 288 active · up to 1M context · 871 thousand downloads in 30 days
GLM-5.3 Flash needs about 201 GB at UD-Q4_K_XL with a 32K-token context. The file itself is 200 GB, and the cache for 32K tokens adds 0.6 GB. No single consumer card holds it. A Mac with 256 GB of unified memory does, and a PC can split it between a GPU and system RAM. It routes each token through 8 of 288 experts, so it runs faster than its size suggests.
Only 11 of its 45 layers keep a growing cache (the rest carry a fixed-size state), so the full 1M context costs 14.9 GB of cache.
| Quant | File | At 8K | At 32K | Smallest GPU | Smallest Mac |
|---|---|---|---|---|---|
UD-IQ1_S | 93.1 GB | 94.2 GB | 94.5 GB | 96 GB | 96 GB* |
UD-Q2_K_XL | 109 GB | 110 GB | 110 GB | none | 128 GB* |
UD-Q3_K_XL | 148 GB | 149 GB | 149 GB | none | 256 GB |
UD-Q4_K_XL | 200 GB | 201 GB | 201 GB | none | 256 GB |
UD-Q5_K_XL | 240 GB | 241 GB | 242 GB | none | 256 GB* |
UD-Q6_K_XL | 292 GB | 293 GB | 293 GB | none | 512 GB |
Q8_0 | 341 GB | 342 GB | 342 GB | none | 512 GB |
BF16 | 642 GB | 643 GB | 643 GB | none | none |
Smallest GPU and Mac are for the 32K total. * with the Mac's GPU memory limit raised.
Sources: GGUF files · config.json
GLM-5.3
754 billion parameters · mixture of experts, 8 of 256 active · up to 1M context · 606 thousand downloads in 30 days
At UD-Q4_K_XL, the GLM-5.3 file is 467 GB, and with 32K tokens of context and runtime overhead it comes to about 472 GB. No single consumer card holds it. A Mac with 512 GB of unified memory does once you raise the GPU memory limit, and a PC can split it between a GPU and system RAM. It routes each token through 8 of 256 experts, so it runs faster than its size suggests.
It caches a compressed latent per token rather than full keys and values, about 107 KB a token, so the cache is small next to the weights: 115 GB at 1M.
| Quant | File | At 8K | At 32K | Smallest GPU | Smallest Mac |
|---|---|---|---|---|---|
UD-IQ1_S | 217 GB | 218 GB | 221 GB | none | 256 GB* |
UD-Q2_K_XL | 254 GB | 256 GB | 258 GB | none | 256 GB* |
UD-Q3_K_XL | 343 GB | 345 GB | 347 GB | none | 512 GB |
UD-Q4_K_XL | 467 GB | 469 GB | 472 GB | none | 512 GB* |
UD-Q5_K_XL | 562 GB | 564 GB | 567 GB | none | none |
UD-Q6_K_XL | 684 GB | 686 GB | 689 GB | none | none |
Q8_0 | 801 GB | 803 GB | 806 GB | none | none |
BF16 | 1,508 GB | 1,510 GB | 1,512 GB | none | none |
Smallest GPU and Mac are for the 32K total. * with the Mac's GPU memory limit raised.
Sources: GGUF files · config.json
GLM-5.2
754 billion parameters · mixture of experts, 8 of 256 active · up to 1M context · 344 thousand downloads in 30 days
GLM-5.2 needs about 470 GB at UD-Q4_K_M with a 32K-token context. The file itself is 466 GB, and the cache for 32K tokens adds 3.6 GB. No single consumer card holds it. A Mac with 512 GB of unified memory does once you raise the GPU memory limit, and a PC can split it between a GPU and system RAM. It routes each token through 8 of 256 experts, so it runs faster than its size suggests.
It caches a compressed latent per token rather than full keys and values, about 107 KB a token, so the cache is small next to the weights: 115 GB at 1M.
| Quant | File | At 8K | At 32K | Smallest GPU | Smallest Mac |
|---|---|---|---|---|---|
UD-IQ1_S | 217 GB | 218 GB | 221 GB | none | 256 GB* |
UD-Q2_K_XL | 254 GB | 256 GB | 258 GB | none | 256 GB* |
UD-Q3_K_XL | 343 GB | 345 GB | 347 GB | none | 512 GB |
UD-Q4_K_M | 466 GB | 468 GB | 470 GB | none | 512 GB* |
UD-Q5_K_XL | 562 GB | 564 GB | 567 GB | none | none |
UD-Q6_K_XL | 684 GB | 686 GB | 689 GB | none | none |
Q8_0 | 801 GB | 803 GB | 806 GB | none | none |
BF16 | 1,508 GB | 1,510 GB | 1,512 GB | none | none |
Smallest GPU and Mac are for the 32K total. * with the Mac's GPU memory limit raised.
Sources: GGUF files · config.json
GLM-4.7 Flash
29.9 billion parameters · mixture of experts, 4 of 64 active · up to 198K context · 141 thousand downloads in 30 days
At Q4_K_M, the GLM-4.7 Flash file is 18.3 GB, and with 32K tokens of context and runtime overhead it comes to about 20.9 GB. The smallest card that holds it whole is a 24 GB card (RTX 3090, RTX 4090). It routes each token through 4 of 64 experts, so it runs faster than its size suggests.
It caches a compressed latent per token rather than full keys and values, about 53 KB a token, so the cache is small next to the weights: 11.0 GB at 198K.
| Quant | File | At 8K | At 32K | Smallest GPU | Smallest Mac |
|---|---|---|---|---|---|
UD-TQ1_0 | 8.3 GB | 9.6 GB | 10.9 GB | 12 GB | 16 GB |
UD-Q2_K_XL | 11.9 GB | 13.1 GB | 14.5 GB | 16 GB | 24 GB |
UD-Q3_K_XL | 13.8 GB | 15.0 GB | 16.4 GB | 16 GB | 24 GB |
Q4_K_M | 18.3 GB | 19.6 GB | 20.9 GB | 24 GB | 24 GB* |
Q5_K_M | 21.4 GB | 22.7 GB | 24.0 GB | 24 GB | 32 GB* |
Q6_K | 24.7 GB | 26.0 GB | 27.3 GB | 32 GB | 32 GB* |
Q8_0 | 31.8 GB | 33.1 GB | 34.4 GB | 48 GB | 48 GB |
BF16 | 59.9 GB | 61.2 GB | 62.5 GB | 96 GB | 64 GB* |
Smallest GPU and Mac are for the 32K total. * with the Mac's GPU memory limit raised.
Sources: GGUF files · config.json
GLM-4.6V Flash
9.4 billion parameters · up to 128K context · 20 thousand downloads in 30 days
GLM-4.6V Flash needs about 8.3 GB at Q4_K_M with a 32K-token context. The file itself is 6.2 GB, and the cache for 32K tokens adds 1.3 GB. That fits a 8 GB card (RTX 4060, RTX 3070). On a Mac, 16 GB of unified memory is enough.
Every token of context adds 40 KB of cache, so the full 128K would put 5.4 GB on top of the weights.
| Quant | File | At 8K | At 32K | Smallest GPU | Smallest Mac |
|---|---|---|---|---|---|
UD-IQ1_S | 3.2 GB | 4.4 GB | 5.4 GB | 8 GB | 16 GB |
UD-Q2_K_XL | 4.2 GB | 5.4 GB | 6.4 GB | 8 GB | 16 GB |
UD-Q3_K_XL | 5.1 GB | 6.3 GB | 7.3 GB | 8 GB | 16 GB |
Q4_K_M | 6.2 GB | 7.3 GB | 8.3 GB | 8 GB | 16 GB |
Q5_K_M | 7.1 GB | 8.2 GB | 9.2 GB | 12 GB | 16 GB |
Q6_K | 8.3 GB | 9.4 GB | 10.4 GB | 12 GB | 16 GB |
Q8_0 | 10.0 GB | 11.1 GB | 12.2 GB | 12 GB | 16 GB* |
BF16 | 18.8 GB | 20.0 GB | 21.0 GB | 24 GB | 24 GB* |
Smallest GPU and Mac are for the 32K total. * with the Mac's GPU memory limit raised.
Sources: GGUF files · config.json
How the numbers are worked out
Total memory is three parts added together. Weights are the exact size of the GGUF file on Hugging Face, all parts summed when a quant is split. The context cache (the KV cache) comes from each model's own config.json:
cache = layers × kv_heads × head_dim × 2 (K and V) × 2 bytes × tokens
Sliding-window layers stop at their window plus llama.cpp's 512-token batch. Linear-attention and Mamba layers keep a fixed state instead, which we add once. Latent-attention models (DeepSeek, Kimi, GLM-5) cache kv_lora_rank + rope dimensions per layer, the way llama.cpp does. Overhead is our allowance: 0.5 GB for the GPU context plus the output buffer, 512 × vocabulary × 4 bytes.
A "24 GB" card holds 24 GiB, which is 25.8 GB in the decimal units file sizes use, and we count it that way. Macs get the share Metal allows by default, two thirds of unified memory up to 32 GB and three quarters above (llama.cpp discussion 2182); CPU-only machines and the DGX Spark keep 4 GB back for the operating system. All of it assumes llama.cpp's defaults: an f16 cache and one conversation at a time. Quantizing the cache to q8_0 halves it.
Files are Unsloth's GGUF uploads, the most downloaded quantizations on Hugging Face, and the list refreshes daily from Hugging Face's API. A model joins when its repo passes 20,000 downloads in 30 days. Last checked September 25, 2026.