GLM-5 · released 2026-08 · 321B parameters (18.0B active per token)
Hybrid: 34 KDA linear-attention layers (64 heads x 128) + 11 NoPE-MLA layers with DSA indexer (MLA cache 512 elements/token/layer). mHC hyper-connections (x4). Card: 320B total / 18B active, natively multimodal. KDA state assumed BF16.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.11 GB | 0.09 GB |
| 8,192 tokens | 0.16 GB | 0.11 GB |
| 32,768 tokens | 0.41 GB | 0.24 GB |
| 131,072 tokens | 1.45 GB | 0.76 GB |
| 1,048,576 tokens | 11.1 GB | 5.57 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 598.5 GB |
| FP8 (W8A8) | 8 | 301.6 GB |
| INT8 (W8A8 / W8A16) | 8 | 301.6 GB |
| NVFP4 | 4.5 | 171.7 GB |
| INT4 AWQ / GPTQ | 4.25 | 162.5 GB |
| GGUF Q8_0 | 8.5 | 318.0 GB |
| GGUF Q6_K | 6.56 | 245.4 GB |
| GGUF Q5_K_M | 5.69 | 212.8 GB |
| GGUF Q4_K_M | 4.89 | 182.9 GB |
| GGUF Q3_K_M | 3.91 | 146.3 GB |
| GGUF Q2_K | 3.35 | 125.3 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× · 182 tok/s | 4× · 63.0 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 154 tok/s | 4× · 42.0 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 106 tok/s | 8× · 45.0 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 2× · 150 tok/s | 2× · 26.7 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 139 tok/s | 4× · 33.4 tok/s |
| L40S 48GBFP8 (W8A8) | 8× · 42.8 tok/s | 16× · 15.2 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 4× · 65.8 tok/s | 4× · 14.7 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 8× · 49.4 tok/s | 8× · 26.3 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 26.8 tok/s | 16× · 17.9 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 |