GLM-4.5 · released 2025-12 · 358B parameters (33.6B active per token)
GLM-4.5 architecture with 200K context.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 1.44 GB | 0.72 GB |
| 8,192 tokens | 2.88 GB | 1.44 GB |
| 32,768 tokens | 11.5 GB | 5.75 GB |
| 131,072 tokens | 46.0 GB | 23.0 GB |
| 202,752 tokens | 71.2 GB | 35.6 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 667.5 GB |
| FP8 (W8A8) | 8 | 336.6 GB |
| INT8 (W8A8 / W8A16) | 8 | 336.6 GB |
| NVFP4 | 4.5 | 191.9 GB |
| INT4 AWQ / GPTQ | 4.25 | 181.5 GB |
| GGUF Q8_0 | 8.5 | 354.6 GB |
| GGUF Q6_K | 6.56 | 273.7 GB |
| GGUF Q5_K_M | 5.69 | 237.4 GB |
| GGUF Q4_K_M | 4.89 | 204.0 GB |
| GGUF Q3_K_M | 3.91 | 163.1 GB |
| GGUF Q2_K | 3.35 | 139.7 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 4× · 102 tok/s | 4× · 36.8 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 88.4 tok/s | 4× · 24.8 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 56.9 tok/s | 8× · 26.0 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 91.3 tok/s | 4× · 26.9 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 79.8 tok/s | 4× · 18.4 tok/s |
| L40S 48GBFP8 (W8A8) | 16× · 13.5 tok/s | 16× · 8.0 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 8× · 25.8 tok/s | 8× · 11.8 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 8× · 25.8 tok/s | 16× · 10.7 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 13.5 tok/s | 16× · 9.0 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |