StepFun Step 3.5

Step 3.5 Flash hardware requirements

Needs at least 2 GPUs at BF16 (Instinct MI355X OAM). For 1,000 tokens/s at 32k context you need 8 x H100 SXM5 80GB at $14.64/hr or 16 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $8.00/hr.

Parameters

199B

Active per token

11.0B

mixture of experts

KV cache per token

180 KB

16-bit cache

Native precision

BF16

Rent Model card

Top options for Step 3.5 Flash

The strongest part each vendor offers for Step 3.5 Flash at BF16 with 32,768 tokens per user, rentable parts first. Each card is one replica, the fewest accelerators that hold the model, and the line it draws on the graph below. Select a card to plot it.

Tokens per second as users grow

Each line is one replica, the fewest accelerators that hold the model, serving more users at once. It climbs while memory bandwidth is the limit, flattens once compute is, and stops where memory runs out at 32,768 tokens per user. Hover a line to bring it forward.

4 x GB200 NVL72 GPU 186GB 2 x Instinct MI355X OAM 4 x Ascend 950DT 8 x Gaudi 2 96GB 4 x TPU v7 192GB 16 x Inferentia2

Memory fills at: 4 x GB200 NVL72 GPU 186GB 188 users, 2 x Instinct MI355X OAM 92 users, 4 x Ascend 950DT 91 users, 8 x Gaudi 2 96GB 199 users, 4 x TPU v7 192GB 201 users, 16 x Inferentia2 46 users. Roofline estimates at BF16.

Bare minimum

Fewest GPUs

2 x Instinct MI355X OAM

BF16, 4,096 tokens, one stream

Tokens/s: 505
Per hour: $11.98
Per million tokens: $6.59
Memory used: 72%

Rent on Daytona

Cheapest per hour

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 4,096 tokens, one stream, live price

Tokens/s: 403
Per hour: $4.00
Per million tokens: $2.76
Memory used: 54%

Rent on RunPod

Cheapest per million tokens

8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition

BF16, 32,768 tokens, 32 streams

Tokens/s: 716
Per hour: $4.00
Per million tokens: $1.55
Memory used: 62%

Rent on RunPod

Configurations

Every datacenter GPU in the catalog, sized for these settings. Change precision, context and concurrency and the whole page follows. Each point is one replica: the smallest tensor-parallel group of that GPU that holds the model.

Precision Context Streams

Vendor

Priced only

Priced configurations only; 55 more fit without a live price and appear in the table. Marker size grows with the tensor-parallel width. Roofline estimates.

#									Rent
1	Instinct MI455X	2	54 %	2.5k tok/s	112 tok/s	—	—	—	Specs →
2	Rubin SXM	2	82 %	2.2k tok/s	106 tok/s	—	—	—	Specs →
3	TPU 8t announced	4	54 %	~1.6k tok/s	59.7 tok/s	—	—	—	Specs →
4	TPU 8i announced	2	82 %	~1.0k tok/s	41.5 tok/s	—	—	—	Specs →
5	Instinct MI355X OAM	2	82 %	952 tok/s	38.6 tok/s	$11.98	$3.50	2.8 kW	Rent →
6	GB300 NVL72 GPU 288GB	2	82 %	950 tok/s	38.6 tok/s	$24.84	$7.26	—	Rent →
7	GB200 NVL72 GPU 186GB	4	63 %	1.8k tok/s	73.2 tok/s	$42.00	$6.48	—	Rent →
8	Instinct MI350X OAM	2	82 %	932 tok/s	38.6 tok/s	$12.32	$3.67	2.0 kW	Rent →
9	TPU v7 192GB	4	61 %	1.7k tok/s	67.5 tok/s	—	—	—	Specs →
10	Instinct MI430X OAM announced	2	54 %	~1.8k tok/s	112 tok/s	—	—	—	Specs →
11	B300 SXM 262GB	2	90 %	926 tok/s	38.6 tok/s	$9.98	$2.99	2.8 kW	Rent →
12	B200 SXM 180GB	4	65 %	1.7k tok/s	70.4 tok/s	$16.36	$2.66	4.0 kW	Rent →
13	Ascend 970 announced	2	82 %	1.3k tok/s	69.5 tok/s	—	—	—	Specs →
14	B100 SXM 192GB	4	61 %	1.6k tok/s	73.2 tok/s	—	—	2.8 kW	Specs →
15	Jalapeño announced	4	54 %	~2.4k tok/s	141 tok/s	—	—	2.8 kW	Specs →
16	Instinct MI325X OAM	2	92 %	648 tok/s	29.0 tok/s	$7.60	$3.26	2.0 kW	Rent →
17	Instinct MI300X 192GB	4	61 %	1.1k tok/s	48.5 tok/s	$9.56	$2.36	3.0 kW	Rent →
18	BR100	8	93 %	771 tok/s	27.6 tok/s	—	—	4.4 kW	Specs →
19	Ascend 960 announced	2	82 %	780 tok/s	46.3 tok/s	—	—	—	Specs →
20	H200 SXM 141GB	4	84 %	965 tok/s	43.9 tok/s	$9.92	$2.85	2.8 kW	Rent →
21	H100 SXM5 80GB	8	74 %	1.4k tok/s	57.9 tok/s	$14.64	$2.89	5.6 kW	Rent →
22	TPU v6e 32GB 2 nodes	16	95 %	1.4k tok/s	49.9 tok/s	—	—	—	Specs →
23	Data Center GPU Max 1550 128GB	4	92 %	702 tok/s	30.0 tok/s	—	—	2.4 kW	Specs →
24	H200 NVL 141GB	4	84 %	~913 tok/s	43.9 tok/s	$14.46	$4.40	2.4 kW	Rent →
25	H100 NVL 94GB	8	63 %	~1.5k tok/s	67.4 tok/s	$25.52	$4.73	3.2 kW	Rent →
26	Ascend 910C	4	92 %	681 tok/s	29.3 tok/s	—	—	2.4 kW	Specs →
27	H100 PCIe 80GB	8	74 %	886 tok/s	34.6 tok/s	$20.00	$6.27	2.8 kW	Rent →
28	H800 PCIe 80GB	8	74 %	886 tok/s	34.6 tok/s	—	—	2.8 kW	Specs →
29	Ascend 950DT	4	82 %	694 tok/s	36.6 tok/s	—	—	—	Specs →
30	RTX PRO 6000 Blackwell Workstation Edition	8	62 %	743 tok/s	31.0 tok/s	$14.40	$5.39	4.8 kW	Rent →
31	Ascend 950PR	4	92 %	360 tok/s	14.6 tok/s	—	—	3.6 kW	Specs →
32	MTT S5000	8	74 %	~680 tok/s	27.6 tok/s	—	—	—	Specs →
33	RTX PRO 6000 Blackwell Server Edition	8	62 %	679 tok/s	27.6 tok/s	$5.52	$2.26	4.8 kW	Rent →
34	Gaudi 3 128GB	4	92 %	897 tok/s	33.8 tok/s	—	—	3.6 kW	Specs →
35	TPU v5p 95GB	8	62 %	976 tok/s	47.8 tok/s	—	—	3.6 kW	Specs →
36	RTX PRO 6000 Blackwell Max-Q Workstation Edition	8	62 %	716 tok/s	31.0 tok/s	$4.00	$1.55	2.4 kW	Rent →
37	Cloud AI 100 Ultra	4	92 %	~143 tok/s	5.0 tok/s	—	—	0.6 kW	Specs →
38	Gaudi 2 96GB	8	62 %	884 tok/s	42.3 tok/s	$8.80	$2.77	4.8 kW	Rent →
39	Atlas 350	8	53 %	591 tok/s	24.2 tok/s	—	—	4.8 kW	Specs →
40	Data Center GPU Max 1100 48GB 2 nodes	16	62 %	940 tok/s	37.5 tok/s	—	—	4.8 kW	Specs →
41	GeForce RTX 5090 32GB 2 nodes	16	95 %	970 tok/s	54.6 tok/s	$19.04	$5.45	9.2 kW	Rent →
42	Ascend 910B	8	93 %	643 tok/s	27.6 tok/s	—	—	3.2 kW	Specs →
43	MI250X 128GB	4	92 %	532 tok/s	30.0 tok/s	—	—	2.2 kW	Specs →
44	RTX 6000 Ada 48GB 2 nodes	16	62 %	~751 tok/s	29.3 tok/s	$15.52	$5.74	4.8 kW	Rent →
45	Instinct MI250 128GB	4	92 %	518 tok/s	30.0 tok/s	—	—	2.0 kW	Specs →
46	L40S 48GB 2 nodes	16	62 %	689 tok/s	26.3 tok/s	$12.16	$4.91	5.6 kW	Rent →
47	A800 PCIe 80GB	8	74 %	676 tok/s	33.4 tok/s	—	—	2.4 kW	Specs →
48	A800 SXM4	8	74 %	699 tok/s	35.2 tok/s	—	—	3.2 kW	Specs →
49	A100 PCIe 80GB	8	74 %	676 tok/s	33.4 tok/s	$10.80	$4.44	2.4 kW	Rent →
50	A100 SXM4 80GB	8	74 %	699 tok/s	35.2 tok/s	$11.20	$4.45	3.2 kW	Rent →
51	A100 PCIe 40GB 2 nodes	16	75 %	1.0k tok/s	47.4 tok/s	—	—	4.0 kW	Specs →
52	A100 SXM4 40GB 2 nodes	16	75 %	1.0k tok/s	47.4 tok/s	$24.98	$6.71	6.4 kW	Rent →
53	RTX 5880 Ada 48GB 2 nodes	16	62 %	~707 tok/s	29.3 tok/s	—	—	4.6 kW	Specs →
54	TPU v4 32GB 2 nodes	16	95 %	830 tok/s	36.6 tok/s	—	—	—	Specs →
55	RTX 5000 Ada 32GB 2 nodes	16	95 %	~466 tok/s	17.6 tok/s	$13.28	$7.92	4.0 kW	Rent →
56	RTX PRO 5000 Blackwell 48GB 2 nodes	16	62 %	~881 tok/s	41.0 tok/s	$15.36	$4.85	4.8 kW	Rent →
57	RTX PRO 5000 Blackwell 72GB	8	83 %	~499 tok/s	23.2 tok/s	—	—	2.4 kW	Specs →
58	RTX PRO 4500 Blackwell Server Edition 2 nodes	16	95 %	570 tok/s	24.4 tok/s	—	—	2.6 kW	Specs →
59	RTX PRO 4500 Blackwell 32GB 2 nodes	16	95 %	618 tok/s	27.3 tok/s	$11.52	$5.18	3.2 kW	Rent →
60	Radeon AI PRO R9700 32GB 2 nodes	16	95 %	475 tok/s	19.5 tok/s	—	—	4.8 kW	Specs →
61	Inferentia2 2 nodes	16	95 %	569 tok/s	25.0 tok/s	—	—	—	Specs →
62	Trainium 2 nodes	16	95 %	569 tok/s	25.0 tok/s	—	—	—	Specs →
63	Instinct MI100 32GB 2 nodes	16	95 %	533 tok/s	37.5 tok/s	—	—	4.8 kW	Specs →
64	Arc Pro B70 32GB 2 nodes	16	95 %	~452 tok/s	18.5 tok/s	—	—	3.7 kW	Specs →
65	L40 48GB 2 nodes	16	62 %	582 tok/s	26.3 tok/s	$13.12	$6.26	4.8 kW	Rent →
66	Instinct MI210 PCIe	8	93 %	490 tok/s	28.3 tok/s	—	—	2.4 kW	Specs →
67	RTX A6000 48GB 2 nodes	16	62 %	511 tok/s	23.4 tok/s	$8.00	$4.35	4.8 kW	Rent →
68	A40 48GB 2 nodes	16	62 %	473 tok/s	21.2 tok/s	$9.44	$5.55	4.8 kW	Rent →
69	H20 141GB HBM3e	4	84 %	331 tok/s	36.6 tok/s	—	—	1.6 kW	Specs →
70	H20 96GB	8	62 %	625 tok/s	69.1 tok/s	—	—	3.2 kW	Specs →
71	BI-V100 Tiangai 100 2 nodes	16	95 %	663 tok/s	36.6 tok/s	—	—	4.0 kW	Specs →
72	Atlas 300I Duo 48GB 2 nodes	16	62 %	313 tok/s	12.4 tok/s	—	—	2.4 kW	Specs →
73	Atlas 300I Duo 96GB	8	62 %	177 tok/s	7.0 tok/s	—	—	1.2 kW	Specs →
74	V100S PCIe 32GB 2 nodes	16	95 %	608 tok/s	34.6 tok/s	$14.08	$6.43	4.0 kW	Rent →
75	ATOM-Max	8	93 %	323 tok/s	17.7 tok/s	—	—	2.8 kW	Specs →
76	Radeon PRO W7900 Dual Slot 48GB 2 nodes	16	62 %	508 tok/s	26.3 tok/s	—	—	4.7 kW	Specs →
77	Radeon PRO W7900 48GB 2 nodes	16	62 %	508 tok/s	26.3 tok/s	—	—	4.7 kW	Specs →
78	L20 48GB 2 nodes	16	62 %	502 tok/s	26.3 tok/s	—	—	4.4 kW	Specs →
79	Radeon AI PRO R9600D 32GB 2 nodes	16	95 %	389 tok/s	19.5 tok/s	—	—	2.4 kW	Specs →
80	Arc Pro B65 32GB 2 nodes	16	95 %	~376 tok/s	18.5 tok/s	—	—	3.2 kW	Specs →
81	MLU370-X8 2 nodes	16	62 %	374 tok/s	18.7 tok/s	—	—	4.0 kW	Specs →
82	Radeon PRO W7800 48GB 2 nodes	16	62 %	444 tok/s	26.3 tok/s	—	—	4.2 kW	Specs →
83	Radeon PRO W7800 32GB 2 nodes	16	95 %	352 tok/s	17.6 tok/s	—	—	4.2 kW	Specs →
84	Atlas 300V Pro 2 nodes	16	62 %	157 tok/s	6.2 tok/s	—	—	—	Specs →
85	Radeon PRO W6800 32GB 2 nodes	16	95 %	213 tok/s	15.6 tok/s	—	—	4.0 kW	Specs →

By vendor

At BF16 the NVIDIA pick is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at about 716 tokens/s for $4.00/hr; the AMD pick is 4 x Instinct MI300X 192GB at about 1.1k tokens/s for $9.56/hr. Per rental dollar NVIDIA delivers 1.52x the tokens of AMD here.

NVIDIA 37 parts fit

Best value: 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition, 716 tok/s for $4.00/hr

Fewest GPUs: 2 x Rubin SXM, 2.2k tok/s

AMD 16 parts fit

Best value: 4 x Instinct MI300X 192GB, 1.1k tok/s for $9.56/hr

Fewest GPUs: 2 x Instinct MI455X, 2.5k tok/s

Huawei 8 parts fit

Fewest GPUs: 4 x Ascend 950DT, 694 tok/s

Intel 6 parts fit

Best value: 8 x Gaudi 2 96GB, 884 tok/s for $8.80/hr

Fewest GPUs: 4 x Gaudi 3 128GB, 897 tok/s

Google 4 parts fit

Fewest GPUs: 4 x TPU v7 192GB, 1.7k tok/s

AWS 2 parts fit

Fewest GPUs: 16 x Inferentia2, 569 tok/s

Other 6 parts fit

Fewest GPUs: 4 x Cloud AI 100 Ultra, 143 tok/s

Fleet what-if

Compare whole fleets at these settings: a thousand of one part against a hundred of another. Each fleet splits into replicas of its tensor-parallel width; GPUs left over sit idle.

A GPUs Part

B GPUs Part

Fleet letter	Fleet	Replicas	Idle	Tokens/s	Per hour	Power	Per M tokens
A	1,000 x H200 SXM 141GB	250 x TP4	0	241k tok/s	$2,480	700 kW	$2.85
B	100 x GB200 NVL72 GPU 186GB	25 x TP4	0	45k tok/s	$1,050	—	$6.48

Size Step 3.5 Flash for a tokens per second target

The calculator starts from your traffic instead of a fleet: set a target rate and read the replica count for every GPU.

Open in the LLM GPU Calculator

Frequently asked questions

How much GPU memory does Step 3.5 Flash need?

The weights take 399 GB at the native BF16 precision (399 GB at BF16, 199 GB at FP8, 112 GB at INT4). Each concurrent stream adds 180 KB of KV cache per token: 0.3 GB at 4,096 tokens and 1.7 GB at 32,768 tokens.

What is the bare minimum to run Step 3.5 Flash?

2 x Instinct MI355X OAM at BF16 with 4,096 tokens of context and one stream. The cheapest live rental for that footprint is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $4.00 per hour.

How many H100s do you need to run Step 3.5 Flash?

8 H100s at BF16 for 4,096 tokens of context and one stream. For 32,768 tokens and 32 concurrent streams the tensor-parallel group is 8 H100s, and 1,000 tokens per second takes 8 H100s across 1 replicas.

How much does Step 3.5 Flash cost per million tokens on an H100?

About $2.89 per million output tokens at BF16, 32,768 tokens of context and 32 streams, using the cheapest live on-demand price of $1.83 per GPU-hour. The cheapest part per token is 8 x RTX PRO 6000 Blackwell Max-Q Workstation Edition at $1.55.

What precision does Step 3.5 Flash ship in?

The published checkpoint is BF16. FP8 and INT4 figures describe post-training quantisations that halve and quarter the footprint at a small accuracy cost.

All figures are roofline estimates from the model's config.json and each GPU's published memory, bandwidth and tensor-core peak, with fixed efficiency factors. Prices are the cheapest live on-demand listing per GPU when the site has one. They are a planning floor; a tuned serving stack can do better. 33 of 45 layers attend over a 512-token sliding window. The checkpoint carries three multi-token-prediction layers (199.4B tensors against the 196.8B the card quotes). Active parameters per the StepFun model card.