Qwen3 · released 2025-04 · 235B parameters (22.2B active per token)
Same architecture for the original (Apr 2025, 32K native/131K YaRN) and the 2507 Instruct/Thinking updates (262144 native, config shown). params = original BF16 checkpoint.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.73 GB | 0.37 GB |
| 8,192 tokens | 1.47 GB | 0.73 GB |
| 32,768 tokens | 5.88 GB | 2.94 GB |
| 131,072 tokens | 23.5 GB | 11.8 GB |
| 262,144 tokens | 47.0 GB | 23.5 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 437.9 GB |
| FP8 (W8A8) | 8 | 221.3 GB |
| INT8 (W8A8 / W8A16) | 8 | 221.3 GB |
| NVFP4 | 4.5 | 126.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 119.7 GB |
| GGUF Q8_0 | 8.5 | 232.6 GB |
| GGUF Q6_K | 6.56 | 179.5 GB |
| GGUF Q5_K_M | 5.69 | 155.7 GB |
| GGUF Q4_K_M | 4.89 | 133.8 GB |
| GGUF Q3_K_M | 3.91 | 107.0 GB |
| GGUF Q2_K | 3.35 | 91.7 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) | 2× · 155 tok/s | 2× · 32.9 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 2× · 122 tok/s | 4× · 33.5 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 4× · 93.3 tok/s | 4× · 25.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 2× · 128 tok/s | 2× · 22.7 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 2× · 104 tok/s | 4× · 25.9 tok/s |
| L40S 48GBFP8 (W8A8) | 8× · 25.0 tok/s | 8× · 9.4 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 4× · 45.2 tok/s | 4× · 11.9 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 8× · 26.2 tok/s | 8× · 15.8 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 8× · 25.1 tok/s | 16× · 9.5 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 |