Qwen3-Coder · released 2025-07 · 480B parameters (35.5B active per token)
262144 native context (1M with YaRN per card).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.97 GB | 0.48 GB |
| 8,192 tokens | 1.94 GB | 0.97 GB |
| 32,768 tokens | 7.75 GB | 3.88 GB |
| 131,072 tokens | 31.0 GB | 15.5 GB |
| 262,144 tokens | 62.0 GB | 31.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 894.4 GB |
| FP8 (W8A8) | 8 | 450.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 450.7 GB |
| NVFP4 | 4.5 | 256.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 242.7 GB |
| GGUF Q8_0 | 8.5 | 475.1 GB |
| GGUF Q6_K | 6.56 | 366.7 GB |
| GGUF Q5_K_M | 5.69 | 318.1 GB |
| GGUF Q4_K_M | 4.89 | 273.3 GB |
| GGUF Q3_K_M | 3.91 | 218.6 GB |
| GGUF Q2_K | 3.35 | 187.3 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× · 136 tok/s | 4× · 33.9 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 114 tok/s | 8× · 32.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 78.6 tok/s | 8× · 25.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 119 tok/s | 4× · 23.8 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 102 tok/s | 8× · 26.5 tok/s |
| L40S 48GBFP8 (W8A8) | 16× · 19.5 tok/s | 16× · 9.0 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 8× · 36.1 tok/s | 8× · 11.8 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 16× · 20.1 tok/s | 16× · 13.9 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 19.6 tok/s | > 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 |