Step-3.5-Flash — which GPUs fit

StepFun · 196.8B (active 11.0B) · MoE · ctx 256K · apache-2.0 · HuggingFace

Step-3.5-Flash (StepFun) is a 196.8B parameters MoE (11.0B active per token) open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 111.44 GB VRAM for Q4 (weights 108.24 GB + KV 2.2 GB + 1 GB overhead) and 396.8 GB for FP16. About 7 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 462.8 tok/s (estimated); MI300X 192GB ≈ 306.6 tok/s (estimated); H200 141GB SXM5 ≈ 277.7 tok/s (estimated); Gaudi 3 128GB ≈ 189.5 tok/s (estimated); Mac Studio M2 Ultra 192GB ≈ 46.3 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.

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

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	462.8 est.	Yes	0.012
MI300X 192GB	192	Q4	306.6 est.	Yes	0.02
H100 80GB SXM5	80	Q4	285.4 est.	No	0.01
H200 141GB SXM5	141	Q4	277.7 est.	Yes	0.008
Gaudi 3 128GB	128	Q4	189.5 est.	Yes	0.015
A100 80GB SXM4	80	Q4	173.7 est.	No	0.012
Mac Studio M3 Ultra 96GB	96	Q4	100.5 est.	No	0.025
B200 192GB SXM	192	FP16	85.1 est.	No	0.002
MI300X 192GB	192	FP16	56.4 est.	No	0.004
Mac Studio M2 Ultra 192GB	192	Q4	46.3 est.	Yes	0.007
MacBook Pro M4 Max 128GB	128	Q4	31.6 est.	Yes	0.007
RTX 6000 Ada 48GB	48	Q4	29.4 est.	No	0.004
H200 141GB SXM5	141	FP16	27.5 est.	No	0.001
L40S 48GB	48	Q4	26.5 est.	No	0.003
RTX 5090 32GB	32	Q4	24.4 est.	No	0.012
RTX A6000 48GB	48	Q4	23.6 est.	No	0.005
MacBook Pro M3 Max 128GB	128	Q4	23.1 est.	Yes	0.006
Gaudi 3 128GB	128	FP16	15.5 est.	No	0.001
Mac mini M4 Pro 64GB	64	Q4	14.9 est.	No	0.007
Mac Studio M2 Ultra 192GB	192	FP16	8.5 est.	No	0.001
MacBook Pro M4 Max 128GB	128	FP16	2.6 est.	No	0.001
MacBook Pro M3 Max 128GB	128	FP16	1.9 est.	No	0.0
Mac mini M4 32GB	32	Q4	1.6 est.	No	0.002

Fit checks

RTX 5090 32GB — does not fit · ≈24.4 tok/s Q4 est.
RTX 6000 Ada 48GB — does not fit · ≈29.4 tok/s Q4 est.
L40S 48GB — does not fit · ≈26.5 tok/s Q4 est.
Mac Studio M3 Ultra 96GB — does not fit · ≈100.5 tok/s Q4 est.
Mac Studio M2 Ultra 192GB — fits · ≈46.3 tok/s Q4 est.

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