← AI Hashrate

Qwen3.5-397B-A17B

Alibaba · 397.0B (active 17.0B) · MoE · ctx 256K · apache-2.0 · HuggingFace

Qwen3.5-397B-A17B (Alibaba) is a 397.0B parameters MoE (17.0B active per token) open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 222.75 GB VRAM for Q4 (weights 218.35 GB + KV 3.4 GB + 1 GB overhead) and 798.4 GB for FP16. About 0 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 635.7 tok/s (estimated); MI300X 192GB ≈ 421.1 tok/s (estimated); H200 141GB SXM5 ≈ 205.7 tok/s (estimated). Use the table below for VRAM fit and tok/s per dollar (list MSRP). Estimates follow memory-bandwidth math; measured rows override when present. See methodology for details. Methodology.

11 hardware configs — sorted by tok/s. Default fit context 8K.

HardwareVRAMQuanttok/sFits?tok/s/$
B200 192GB SXM192Q4635.7 est.No0.016
MI300X 192GB192Q4421.1 est.No0.028
H200 141GB SXM5141Q4205.7 est.No0.006
Gaudi 3 128GB128Q4115.7 est.No0.009
Mac Studio M2 Ultra 192GB192Q463.6 est.No0.009
H100 80GB SXM580Q446.2 est.No0.002
A100 80GB SXM480Q428.1 est.No0.002
MacBook Pro M4 Max 128GB128Q419.3 est.No0.004
Mac Studio M3 Ultra 96GB96Q416.3 est.No0.004
MacBook Pro M3 Max 128GB128Q414.1 est.No0.003
Mac mini M4 Pro 64GB64Q42.4 est.No0.001

Fit checks

12 popular retail GPUs; every row in the table above also links its Fits? verdict to the full fit check.