Qwen3.5 · released 2026-02 · 9.65B parameters
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 | 0.17 GB | 0.11 GB |
| 8,192 tokens | 0.30 GB | 0.17 GB |
| 32,768 tokens | 1.05 GB | 0.55 GB |
| 131,072 tokens | 4.05 GB | 2.05 GB |
| 262,144 tokens | 8.05 GB | 4.05 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 18.0 GB |
| FP8 (W8A8) | 8 | 12.8 GB |
| INT8 (W8A8 / W8A16) | 8 | 12.8 GB |
| NVFP4 | 4.5 | 10.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 10.3 GB |
| GGUF Q8_0 | 8.5 | 9.55 GB |
| GGUF Q6_K | 6.56 | 7.37 GB |
| GGUF Q5_K_M | 5.69 | 6.39 GB |
| GGUF Q4_K_M | 4.89 | 5.50 GB |
| GGUF Q3_K_M | 3.91 | 4.39 GB |
| GGUF Q2_K | 3.35 | 3.76 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× · 230 tok/s | 1× · 177 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 155 tok/s | 1× · 116 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 115 tok/s | 1× · 84.3 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 168 tok/s | 1× · 126 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 88.9 tok/s | 1× · 61.0 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 32.8 tok/s | 1× · 23.5 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 58.8 tok/s | 1× · 42.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 108 tok/s | 1× · 62.6 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 68.7 tok/s | 1× · 37.7 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 47.5 tok/s | 1× · 11.0 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 10.7 tok/s | 1× · 7.6 tok/s |