Qwen3.5 · released 2026-02 · 2.27B 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). Vision encoder (~0.33B by difference) is ~15% of total at this size.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.07 GB | 0.04 GB |
| 8,192 tokens | 0.11 GB | 0.07 GB |
| 32,768 tokens | 0.39 GB | 0.21 GB |
| 131,072 tokens | 1.52 GB | 0.77 GB |
| 262,144 tokens | 3.02 GB | 1.52 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 4.24 GB |
| FP8 (W8A8) | 8 | 2.59 GB |
| INT8 (W8A8 / W8A16) | 8 | 2.59 GB |
| NVFP4 | 4.5 | 1.87 GB |
| INT4 AWQ / GPTQ | 4.25 | 1.82 GB |
| GGUF Q8_0 | 8.5 | 2.25 GB |
| GGUF Q6_K | 6.56 | 1.74 GB |
| GGUF Q5_K_M | 5.69 | 1.51 GB |
| GGUF Q4_K_M | 4.89 | 1.29 GB |
| GGUF Q3_K_M | 3.91 | 1.04 GB |
| GGUF Q2_K | 3.35 | 0.89 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× · 576 tok/s | 1× · 449 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 461 tok/s | 1× · 335 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 379 tok/s | 1× · 262 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 484 tok/s | 1× · 356 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 346 tok/s | 1× · 207 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 141 tok/s | 1× · 85.9 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 231 tok/s | 1× · 147 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 260 tok/s | 1× · 157 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 195 tok/s | 1× · 104 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 150 tok/s | 1× · 42.1 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 49.8 tok/s | 1× · 29.0 tok/s |