Qwen3 · released 2025-04 · 4.02B parameters
32K native context (max_position_embeddings 40960); 131072 with YaRN per model card. Qwen3-4B-Instruct/Thinking-2507 share this architecture with 262144 native context.
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 | 7.49 GB |
| FP8 (W8A8) | 8 | 4.11 GB |
| INT8 (W8A8 / W8A16) | 8 | 4.11 GB |
| NVFP4 | 4.5 | 2.63 GB |
| INT4 AWQ / GPTQ | 4.25 | 2.52 GB |
| GGUF Q8_0 | 8.5 | 3.98 GB |
| GGUF Q6_K | 6.56 | 3.07 GB |
| GGUF Q5_K_M | 5.69 | 2.66 GB |
| GGUF Q4_K_M | 4.89 | 2.29 GB |
| GGUF Q3_K_M | 3.91 | 1.83 GB |
| GGUF Q2_K | 3.35 | 1.57 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× · 422 tok/s | 1× · 151 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 319 tok/s | 1× · 97.8 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 253 tok/s | 1× · 70.8 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 339 tok/s | 1× · 107 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 239 tok/s | 1× · 47.9 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 84.9 tok/s | 1× · 19.5 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 144 tok/s | 1× · 35.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 166 tok/s | 1× · 35.4 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 115 tok/s | 2× · 36.1 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 83.0 tok/s | 1× · 13.7 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 28.9 tok/s | 1× · 6.3 tok/s |