Qwen3-Coder · released 2025-07 · 30.5B parameters (3.35B 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 | 0.38 GB | 0.19 GB |
| 8,192 tokens | 0.75 GB | 0.38 GB |
| 32,768 tokens | 3.00 GB | 1.50 GB |
| 131,072 tokens | 12.0 GB | 6.00 GB |
| 262,144 tokens | 24.0 GB | 12.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 56.9 GB |
| FP8 (W8A8) | 8 | 29.6 GB |
| INT8 (W8A8 / W8A16) | 8 | 29.6 GB |
| NVFP4 | 4.5 | 17.7 GB |
| INT4 AWQ / GPTQ | 4.25 | 16.8 GB |
| GGUF Q8_0 | 8.5 | 30.2 GB |
| GGUF Q6_K | 6.56 | 23.3 GB |
| GGUF Q5_K_M | 5.69 | 20.2 GB |
| GGUF Q4_K_M | 4.89 | 17.4 GB |
| GGUF Q3_K_M | 3.91 | 13.9 GB |
| GGUF Q2_K | 3.35 | 11.9 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) | 1× · 442 tok/s | 1× · 93.8 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 351 tok/s | 1× · 59.4 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 287 tok/s | 1× · 42.5 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 369 tok/s | 1× · 65.0 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 286 tok/s | 1× · 37.0 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 105 tok/s | 2× · 21.2 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 172 tok/s | 1× · 20.9 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 175 tok/s | 2× · 44.7 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 127 tok/s | 2× · 28.0 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 95.2 tok/s | 1× · 10.5 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 36.6 tok/s | 1× · 3.7 tok/s |