DeepSeek-V3.2 — which GPUs fit

DeepSeek · 685.4B (active 37B) · MoE · ctx 128K · MIT · HuggingFace

DeepSeek-V3.2 (DeepSeek) is a 685.4B parameters MoE (37B active per token) open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 385.37 GB VRAM for Q4 (weights 376.97 GB + KV 7.4 GB + 1 GB overhead) and 1379.2 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 ≈ 97.6 tok/s (estimated); MI300X 192GB ≈ 64.6 tok/s (estimated); H200 141GB SXM5 ≈ 31.6 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.

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

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	97.6 est.	No	0.002
MI300X 192GB	192	Q4	64.6 est.	No	0.004
H200 141GB SXM5	141	Q4	31.6 est.	No	0.001
Gaudi 3 128GB	128	Q4	17.8 est.	No	0.001
Mac Studio M2 Ultra 192GB	192	Q4	9.8 est.	No	0.001
MacBook Pro M4 Max 128GB	128	Q4	3.0 est.	No	0.001
MacBook Pro M3 Max 128GB	128	Q4	2.2 est.	No	0.001