Qwen3.6 · released 2026-04 · 36.0B parameters (3.45B active per token)
Same architecture as Qwen3.5-35B-A3B. Hybrid 3:1 Gated DeltaNet : gated full attention (head_dim 256, partial RoPE). Natively multimodal; params include vision encoder and 1 MTP layer. DeltaNet state FP32 per config (mamba_ssm_dtype).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.14 GB | 0.10 GB |
| 8,192 tokens | 0.22 GB | 0.14 GB |
| 32,768 tokens | 0.68 GB | 0.37 GB |
| 131,072 tokens | 2.56 GB | 1.31 GB |
| 262,144 tokens | 5.06 GB | 2.56 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 67.0 GB |
| FP8 (W8A8) | 8 | 35.4 GB |
| INT8 (W8A8 / W8A16) | 8 | 35.4 GB |
| NVFP4 | 4.5 | 21.6 GB |
| INT4 AWQ / GPTQ | 4.25 | 20.6 GB |
| GGUF Q8_0 | 8.5 | 35.6 GB |
| GGUF Q6_K | 6.56 | 27.5 GB |
| GGUF Q5_K_M | 5.69 | 23.8 GB |
| GGUF Q4_K_M | 4.89 | 20.5 GB |
| GGUF Q3_K_M | 3.91 | 16.4 GB |
| GGUF Q2_K | 3.35 | 14.0 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× · 408 tok/s | 1× · 122 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 309 tok/s | 1× · 78.1 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 245 tok/s | 1× · 56.2 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 328 tok/s | 1× · 85.4 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 249 tok/s | 1× · 53.1 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 82.3 tok/s | 2× · 28.0 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 140 tok/s | 1× · 27.9 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 164 tok/s | 1× · 40.4 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 2× · 130 tok/s | 2× · 40.4 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 83.2 tok/s | 1× · 15.9 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 28.0 tok/s | 1× · 4.9 tok/s |