Estimates from memory-bandwidth formula; rows marked "measured" override. Methodology.
Fit & speed by quant and context
Quant
Context
VRAM needed
Fits?
tok/s (decode)
Q4
4K
219.56 GB
Yes
18.7±est.
Q4
8K default
219.78 GB
Yes
18.7±est.
Q4
32K
221.05 GB
Yes
18.7±est.
FP16
4K
795.21 GB
No
5.4±est.
FP16
8K default
795.42 GB
No
5.4±est.
FP16
32K
796.7 GB
No
5.3±est.
Fits = weights + KV(ctx) + 1 GB overhead ≤ 95% of VRAM. Measured anchors are context-agnostic; the fit verdict is recomputed per context. A missing tok/s means the model is far beyond this card (offload-only territory).
VRAM breakdown at 8K context
Quant
Weights
KV cache
Overhead
Total needed
DGX Spark (GH200 480GB) VRAM
Q4
218.35 GB
0.43 GB
1 GB
219.78 GB
480 GB
FP16
794.0 GB
0.43 GB
1 GB
795.42 GB
480 GB
Other GPUs that run Qwen3.5-397B-A17B
No catalog GPU fully fits Qwen3.5-397B-A17B at Q4/8K — see the model page for offload estimates. model page