Qwen3 · released 2025-04 · 8.19B parameters
32K native context (max_position_embeddings 40960); 131072 with YaRN per model card.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.56 GB | 0.28 GB |
| 8,192 tokens | 1.13 GB | 0.56 GB |
| 32,768 tokens | 4.50 GB | 2.25 GB |
| 131,072 tokens | 18.0 GB | 9.00 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 15.3 GB |
| FP8 (W8A8) | 8 | 9.95 GB |
| INT8 (W8A8 / W8A16) | 8 | 9.95 GB |
| NVFP4 | 4.5 | 7.62 GB |
| INT4 AWQ / GPTQ | 4.25 | 7.46 GB |
| GGUF Q8_0 | 8.5 | 8.11 GB |
| GGUF Q6_K | 6.56 | 6.26 GB |
| GGUF Q5_K_M | 5.69 | 5.43 GB |
| GGUF Q4_K_M | 4.89 | 4.66 GB |
| GGUF Q3_K_M | 3.91 | 3.73 GB |
| GGUF Q2_K | 3.35 | 3.19 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× · 263 tok/s | 1× · 124 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 182 tok/s | 1× · 79.4 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 136 tok/s | 1× · 57.1 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 196 tok/s | 1× · 86.8 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 112 tok/s | 1× · 39.1 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 40.0 tok/s | 1× · 15.5 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 71.0 tok/s | 1× · 28.3 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 113 tok/s | 1× · 32.2 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 72.5 tok/s | 2× · 32.7 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 50.4 tok/s | 1× · 10.5 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 13.1 tok/s | 1× · 5.0 tok/s |