Hardware to run StepFun Step 3.5 Flash 197B-A11B (MoE)

May 2026 release. 197 B / 11 B active MoE — #5 on OpenRouter weekly tokens (~2.73 T) per MACGPU May 2026 analysis. Apache 2.0, 256 K context, MTP-3 speculative-decoding head, 288 experts top-8 + 1 shared. Free tier on OpenRouter driving wide adoption.

Quantization Q2 Q2_K Q4 Q4_K_M Q5 Q5_K_M Q8 Q8_0

Availability ⚡ Only plug & play

Cheapest

Quad AMD MI50 32 GB (128 GB) homelab build

AMD · rack/large tower

$2,500

tokens / secQ4

70B 35 t/s

120B-MoE 45 t/s

235B-MoE —

Memory128 GB · 122 usable

Bandwidth1024 GB/s

Idle / Active75 W / 1200 W

Sticker$2,500

Why: Lowest sticker that still fits StepFun Step 3.5 Flash 197B-A11B (MoE) ($2.5k USD).

AmazoneBay (build 4× used)Alibaba

📺 Reviews on YouTube

▶ AMD MI50 32 GB for local AI — Qwen 3.6 & Gemma 4 on llama.cpp / vLLM ▶ Scalable local AI build — Qwen3.6-27B, Gemma4-31B, Qwen3.5-122B

Fastest

DGX B200 — 8× B200 server (1.44 TB HBM3e)

NVIDIA · 10U DGX server

$475,000

tokens / secQ4

70B 180 t/s

120B-MoE 330 t/s

235B-MoE 165 t/s

Memory1440 GB · 1404 usable

Bandwidth8000 GB/s

Idle / Active900 W / 10200 W

Sticker$475,000

Why: Highest measured tg/s — 150 t/s on StepFun Step 3.5 Flash 197B-A11B (MoE)-class models at Q4.

NVIDIA DGX B200SuperMicro HGX B200Lambda Labs

📺 Reviews on YouTube

▶ Inside a 1.44 TB HBM3e NVIDIA HGX B200 AI Server from ASRock Rack

All-rounder

Single AMD Instinct MI355X 288 GB workstation

AMD · 4U server (OAM, liquid-cooled)

$28,000

tokens / secQ4

70B —

120B-MoE 320 t/s

235B-MoE 130 t/s

Memory288 GB · 282 usable

Bandwidth8000 GB/s

Idle / Active140 W / 1400 W

Sticker$28,000

Why: Top quartile across speed, value, memory headroom, and efficiency — the "buy this if unsure" pick.

Dell XE9712AMD MI350 seriesSupermicro UBB

📺 Reviews on YouTube

▶ AMD's MI350/355X Advancing AI Event Recap ▶ AMD AI Event w/ Craft Computing — MI350/355X Launch, ROCm 7 ▶ AMD Visit and Tour! Featuring ROCm 7 and AMD Instinct

Best value

Single AMD Instinct MI325X 256 GB workstation

AMD · workstation / 4U server (OAM)

$25,000

tokens / secQ4

70B —

120B-MoE 260 t/s

235B-MoE 105 t/s

Memory256 GB · 250 usable

Bandwidth6000 GB/s

Idle / Active100 W / 1000 W

Sticker$25,000

Why: Best $/tg-per-second — ~$212 per t/s.

AMD directSupermicro UBBTensorWave (cloud)

📺 Reviews on YouTube

▶ Inside the AMD Instinct MI325X — AI / HPC deep dive

Best CUDA

DGX H200 — 8× H200 server (1.13 TB HBM3e)

NVIDIA · 8U DGX / HGX server rack

$380,000

tokens / secQ4

70B 170 t/s

120B-MoE 310 t/s

235B-MoE 155 t/s

Memory1128 GB · 1100 usable

Bandwidth4800 GB/s

Idle / Active700 W / 6500 W

Sticker$380,000

Why: Strongest CUDA-only software stack among fitting builds.

NVIDIA DGX H200SuperMicro HGXLambda Labs

📺 Reviews on YouTube

▶ Inside the SUPER NVIDIA H200 Server From Supermicro (8U HGX H200)

Most VRAM

12× RTX Pro 6000 Blackwell rack (1152 GB)

NVIDIA · 8U server rack (multi-node, 1-2 chassis)

$118,000

tokens / secQ4

70B 170 t/s

120B-MoE —

235B-MoE 260 t/s

Memory1152 GB · 1116 usable

Bandwidth1792 GB/s

Idle / Active340 W / 7400 W

Sticker$118,000

Why: 1116 GB usable — most headroom for batching and longer contexts.

AmazonCoreWeaveSuperMicroSHI enterprise

Efficient

Mac Studio M3 Ultra 256 GB

Apple · small desktop

$7,999

tokens / secQ4

70B 18 t/s

120B-MoE 55 t/s

235B-MoE 22 t/s

Memory256 GB · 232 usable

Bandwidth819 GB/s

Idle / Active10 W / 180 W

Sticker$7,999

Why: 180 W active — lowest power draw of the fitting builds.

AppleB&H Photo

📺 Reviews on YouTube

▶ M3 Ultra Mac Studio Review

Cheapest

Mac Studio M3 Ultra 256 GB

Apple · small desktop

$7,999

tokens / secQ4

70B 18 t/s

120B-MoE 55 t/s

235B-MoE 22 t/s

Memory256 GB · 232 usable

Bandwidth819 GB/s

Idle / Active10 W / 180 W

Sticker$7,999

Why: Lowest sticker that still fits StepFun Step 3.5 Flash 197B-A11B (MoE) ($8k USD).

AppleB&H Photo

📺 Reviews on YouTube

▶ M3 Ultra Mac Studio Review

Fastest

Single AMD Instinct MI355X 288 GB workstation

AMD · 4U server (OAM, liquid-cooled)

$28,000

tokens / secQ4

70B —

120B-MoE 320 t/s

235B-MoE 130 t/s

Memory288 GB · 282 usable

Bandwidth8000 GB/s

Idle / Active140 W / 1400 W

Sticker$28,000

Why: Highest measured tg/s — 145 t/s on StepFun Step 3.5 Flash 197B-A11B (MoE)-class models at Q4.

Dell XE9712AMD MI350 seriesSupermicro UBB

📺 Reviews on YouTube

▶ AMD's MI350/355X Advancing AI Event Recap ▶ AMD AI Event w/ Craft Computing — MI350/355X Launch, ROCm 7 ▶ AMD Visit and Tour! Featuring ROCm 7 and AMD Instinct

All-rounder

Single AMD Instinct MI325X 256 GB workstation

AMD · workstation / 4U server (OAM)

$25,000

tokens / secQ4

70B —

120B-MoE 260 t/s

235B-MoE 105 t/s

Memory256 GB · 250 usable

Bandwidth6000 GB/s

Idle / Active100 W / 1000 W

Sticker$25,000

Why: Top quartile across speed, value, memory headroom, and efficiency — the "buy this if unsure" pick.

AMD directSupermicro UBBTensorWave (cloud)

📺 Reviews on YouTube

▶ Inside the AMD Instinct MI325X — AI / HPC deep dive

Best value

Single AMD Instinct MI300X 192 GB workstation

AMD · workstation

$30,000

tokens / secQ4

70B —

120B-MoE 200 t/s

235B-MoE 80 t/s

Memory192 GB · 188 usable

Bandwidth5300 GB/s

Idle / Active90 W / 750 W

Sticker$30,000

Why: Best $/tg-per-second — ~$330 per t/s.

ASRock RackHot Aisle (cloud)AMD direct

📺 Reviews on YouTube

▶ AMD MI300X server review 8x GPUs | Llama 405b model tested

Best CUDA

Tinybox Green v2 (4× RTX 5090, 128 GB)

tinycorp · 12U pedestal

$45,000

tokens / secQ4

70B 60 t/s

120B-MoE 160 t/s

235B-MoE 70 t/s

Memory128 GB · 124 usable

Bandwidth1792 GB/s

Idle / Active130 W / 2300 W

Sticker$45,000

Why: Strongest CUDA-only software stack among fitting builds.

tinygrad

📺 Reviews on YouTube

▶ George Hotz — tinygrad / Llama 70B on a tinybox

Most VRAM

Mac Studio M3 Ultra 512 GB

Apple · small desktop

$14,199

tokens / secQ4

70B 18 t/s

120B-MoE 55 t/s

235B-MoE 22 t/s

Memory512 GB · 480 usable

Bandwidth819 GB/s

Idle / Active12 W / 220 W

Sticker$14,199

Why: 480 GB usable — most headroom for batching and longer contexts.

Apple RefurbB&H Photo (used)eBay (used)

📺 Reviews on YouTube

▶ M3 Ultra Mac Studio Review ▶ M3 Ultra vs RTX 5090 — The Final Battle

Efficient

Single H200 141 GB workstation

NVIDIA · workstation / 2U server

$40,000

tokens / secQ4

70B 70 t/s

120B-MoE 112 t/s

235B-MoE 55 t/s

Memory141 GB · 138 usable

Bandwidth4800 GB/s

Idle / Active60 W / 700 W

Sticker$40,000

Why: 700 W active — lowest power draw of the fitting builds.

NVIDIA H200 partnersSHI enterpriseLambda Labs

📺 Reviews on YouTube

▶ Nvidia Wouldn't Send Me This $30,000 GPU — H200 Holy $H!T ▶ Inside the SUPER NVIDIA H200 Server From Supermicro

Every other build that runs StepFun Step 3.5 Flash 197B-A11B (MoE)

17 additional builds fit StepFun Step 3.5 Flash 197B-A11B (MoE) at Q4_K_M (120 GB usable minimum), sorted by sticker price.

Build	Price	Memory	Bandwidth	tg/s (Q4)	Active W	5-yr power
Quad Intel Arc Pro B70 buildIntel · rack/large tower	$6k	128 / 124 GB	608 GB/s	23 t/s	700 W	$2.5k
2× Strix Halo cluster (256 GB unified)AMD · mini-PC pair	$7k	256 / 192 GB	256 GB/s	18 t/s	240 W	$841
2× DGX Spark cluster (256 GB unified, CUDA)NVIDIA · two desktops, 200 G interconnect	$10k	256 / 240 GB	273 GB/s	41 t/s	460 W	$1.7k
Octuple Intel Arc Pro B70 clusterIntel · rack/large tower	$11k	256 / 248 GB	608 GB/s	27 t/s	1450 W	$5k
4× Strix Halo cluster (512 GB unified)AMD · rack of 4 mini-PCs, 10 GbE fabric	$14k	512 / 384 GB	256 GB/s	25 t/s	480 W	$1.7k
Tinybox Red (6× 7900 XTX, 144 GB)tinycorp · 12U pedestal	$15k	144 / 138 GB	960 GB/s	36 t/s	1500 W	$5k
4× DGX Spark cluster (512 GB unified, CUDA)NVIDIA · rack of 4 desktops	$20k	512 / 488 GB	273 GB/s	52 t/s	920 W	$3.4k
Dual RTX Pro 6000 Blackwell buildNVIDIA · workstation	$24k	192 / 188 GB	1792 GB/s	—	1100 W	$3.8k
Tinybox Green (6× RTX 4090, 144 GB)tinycorp · 12U pedestal	$25k	144 / 138 GB	1008 GB/s	50 t/s	2200 W	$8k
8× Strix Halo cluster (1024 GB unified)AMD · rack of 8 mini-PCs, 10/25 GbE fabric	$27k	1024 / 768 GB	256 GB/s	32 t/s	960 W	$3.4k
2× Mac Studio M3 Ultra 512 GB cluster (TB5 / MLX)Apple · two desktops, Thunderbolt 5 RDMA	$28k	1024 / 960 GB	819 GB/s	33 t/s	440 W	$1.5k
Quad RTX Pro 6000 Blackwell build (384 GB)NVIDIA · workstation / 4U pedestal	$38k	384 / 372 GB	1792 GB/s	—	2200 W	$8k
Tinybox Pro (8× RTX 4090, 192 GB)tinycorp · 12U pedestal	$40k	192 / 184 GB	1008 GB/s	64 t/s	3200 W	$11k
8× DGX Spark cluster (1024 GB unified, CUDA)NVIDIA · rack of 8 desktops, 200 GbE fabric	$44k	1024 / 976 GB	273 GB/s	68 t/s	1840 W	$7k
Single B200 180 GB workstationNVIDIA · workstation / 4U server	$47k	180 / 176 GB	8000 GB/s	55 t/s	1000 W	$3.6k
8× RTX Pro 6000 Blackwell server (768 GB)NVIDIA · 4U server (e.g. SuperMicro AS-4125GS-TNRT)	$78k	768 / 744 GB	1792 GB/s	—	4800 W	$16k
8× H100 80 GB serverNVIDIA · server rack	$280k	640 / 620 GB	3350 GB/s	100 t/s	5600 W	$20k

Last updated 2026-07-11