Qwen3.5 · released 2026-02 · 0.87B 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 (~0.10B by difference) is ~11% 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 | 1.63 GB |
| FP8 (W8A8) | 8 | 1.05 GB |
| INT8 (W8A8 / W8A16) | 8 | 1.05 GB |
| NVFP4 | 4.5 | 0.80 GB |
| INT4 AWQ / GPTQ | 4.25 | 0.78 GB |
| GGUF Q8_0 | 8.5 | 0.86 GB |
| GGUF Q6_K | 6.56 | 0.67 GB |
| GGUF Q5_K_M | 5.69 | 0.58 GB |
| GGUF Q4_K_M | 4.89 | 0.50 GB |
| GGUF Q3_K_M | 3.91 | 0.40 GB |
| GGUF Q2_K | 3.35 | 0.34 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× · 735 tok/s | 1× · 540 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 649 tok/s | 1× · 423 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 576 tok/s | 1× · 343 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 667 tok/s | 1× · 446 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 527 tok/s | 1× · 261 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 277 tok/s | 1× · 123 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 408 tok/s | 1× · 203 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 345 tok/s | 1× · 184 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 292 tok/s | 1× · 126 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 246 tok/s | 1× · 86.5 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 110 tok/s | 1× · 42.6 tok/s |