GLM-5 · released 2026-02 · 754B parameters (40.0B active per token)
MLA + DeepSeek Sparse Attention (indexer 32 heads x 128, top-2048; indexer key cached per token per layer). Card: 744B total / 40B active (safetensors 754B incl. MTP). GLM-5.1 (Apr 2026) has the identical config.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.34 GB | 0.17 GB |
| 8,192 tokens | 0.69 GB | 0.34 GB |
| 32,768 tokens | 2.74 GB | 1.37 GB |
| 131,072 tokens | 11.0 GB | 5.48 GB |
| 202,752 tokens | 17.0 GB | 8.48 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 1,404.2 GB |
| FP8 (W8A8) | 8 | 705.6 GB |
| INT8 (W8A8 / W8A16) | 8 | 705.6 GB |
| NVFP4 | 4.5 | 400.0 GB |
| INT4 AWQ / GPTQ | 4.25 | 378.2 GB |
| GGUF Q8_0 | 8.5 | 746.0 GB |
| GGUF Q6_K | 6.56 | 575.7 GB |
| GGUF Q5_K_M | 5.69 | 499.4 GB |
| GGUF Q4_K_M | 4.89 | 429.2 GB |
| GGUF Q3_K_M | 3.91 | 343.1 GB |
| GGUF Q2_K | 3.35 | 294.0 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) | 8× · 70.3 tok/s | 8× · 33.9 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 8× · 65.5 tok/s | 8× · 24.5 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 16× · 16.7 tok/s | 16× · 11.9 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 8× · 66.6 tok/s | 8× · 26.2 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 8× · 62.1 tok/s | 8× · 18.9 tok/s |
| L40S 48GBFP8 (W8A8) | > 16 | > 16 |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 16× · 16.1 tok/s | 16× · 9.1 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | > 16 | > 16 |
| GeForce RTX 4090 24GBGGUF Q4_K_M | > 16 | > 16 |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |