Single GPU · DIY build

1x RTX 5090

The fastest single consumer card: 32 GB of GDDR7 at 1.79 TB/s. It fits 32B dense models at 8-bit or 70B at about 3 bits.

Memory for models

32 GB

VRAM

Memory bandwidth

1,792 GB/s

Parts

~$3,234 + unpriced parts

GPU price not included

Largest at 4-bit

Nemotron 3.5 Lightning 30B-A3B

25.3 GB at 8,192 tokens of context

gpt-oss-20b, 4-bit

290 tok/smedian of 4 runs

Parts

Part	Qty	Each
GPUNVIDIA GeForce RTX 5090	1	not priced
CPUAMD Ryzen 9 9950X (16-core, 170 W)	1	~$649
MotherboardASUS TUF GAMING X870E-PLUS WIFI7	1	~$325
RAMCorsair Vengeance 64GB (2x32GB) DDR5-6000 CL30	1	~$1,275
PSUCorsair RM1000x (1000 W)	1	~$210
CaseCorsair 7000D AIRFLOW full tower (450 mm GPU clearance, 8 slots)	1	~$290
StorageCorsair MP600 CORE XT 2TB PCIe 4.0 NVMe	1	~$485
Priced parts only		$3,234GPU price not included

New parts are priced at the maker’s own store or list price. Used cards have no price here yet, and a total with a missing GPU would understate the build, so none is shown.

What it runs at 4-bit

Every model in the catalog at its 4-bit file with 8,192 tokens of context. 20 of 41 fit. Speeds are the median of measured runs on the same GPUs where any exist; otherwise an estimate from memory bandwidth. Try other quants and context lengths.

Model	Needs	Fit	Generation speed
Kimi K32.8T, 104B active · UD-Q4_K_XL	1406.8 GBcontext memory estimated	Does not fit
DeepSeek V4 Pro1.6T, 49B active · Q4_K_M	890.5 GBcontext memory not counted	Does not fit
Kimi K2.61T, 32B active · UD-Q4_K_XL	545.7 GBcontext memory estimated	Does not fit
DeepSeek V4.1 Flash552B, 16B active · Q4 (DwarfStar)	484.5 GB	Does not fit
GLM-5.3753.3B · UD-Q4_K_XL	437.4 GBcontext memory estimated	Does not fit
DeepSeek V3.2671B, 37B active · Q4_K_M	379.6 GBcontext memory estimated	Does not fit
MiniMax M3428B, 23B active · Q4_K_M	244.8 GBcontext memory not counted	Does not fit
Qwen3.5 397B-A17B397B, 17B active · Q4_K_M	229.1 GBcontext memory estimated	Does not fit
GLM-5.3 Flash320B, 18B active · UD-Q4_K_XL	187.6 GBcontext memory estimated	Does not fit
Hy3295B, 21B active · Q4_K_M	173.6 GB	Does not fit
MiniMax M2.7228.7B · Q4_K_M	132.7 GB	Does not fit
Qwen3.8 Flash-Next125B, 6B active · Q4_K_M	113.1 GBcontext memory estimated	Does not fit
Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M	82.6 GBcontext memory estimated	Does not fit
Qwen3.5 122B-A10B122B, 10B active · Q4_K_M	73.0 GBcontext memory estimated	Does not fit
Mistral Medium 3.5 128B128B · Q4_K_M	74.0 GB	Does not fit
GLM-4.5-Air106B, 12B active · Q4_K_M	70.9 GB	Does not fit
Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M	69.3 GBcontext memory estimated	Does not fit
Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M	62.4 GBcontext memory not counted	Does not fit
gpt-oss-120b117B, 5.1B active · MXFP4 (native)	60.2 GBcontext memory estimated	Does not fit
Qwen3 Coder Next80B, 3B active · Q4_K_M	46.9 GBcontext memory estimated	Does not fit
Llama 3.3 70B Instruct70B · Q4_K_M	43.6 GB	Does not fit
Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M	25.3 GBcontext memory estimated	Fits	~175–256 tok/sestimate
Qwen3.6 35B-A3B35B, 3B active · Q4_K_M	22.4 GBcontext memory estimated	Fits	117 tok/smedian of 1 run
Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M	21.9 GBcontext memory estimated	Fits	~240–350 tok/sestimate
Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M	19.5 GB	Fits	~222–324 tok/sestimate
Gemma 4 31B IT30.7B · Q4_K_M	18.9 GBcontext memory estimated	Fits	61 tok/smedian of 1 run
Granite 4.2 30B29.3B · Q4_K_M	20.3 GB	Fits	~49–77 tok/sestimate
Qwen3.8 27B27B · Q4_K_M	18.2 GBcontext memory estimated	Fits	~50–79 tok/sestimate
Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M	17.4 GBcontext memory estimated	Fits	122 tok/smedian of 2 runs
Qwen3.6 27B27B · Q4_K_M	17.7 GBcontext memory estimated	Fits	~52–82 tok/sestimate
Devstral Small 2 24B (2512)24B · Q4_K_M	16.1 GB	Fits	~61–96 tok/sestimate
gpt-oss-20b21B, 3.6B active · MXFP4 (native)	12.5 GBcontext memory estimated	Fits	290 tok/smedian of 4 runs
Phi-414.7B · Q4_K_M	11.3 GB	Fits	~99–155 tok/sestimate
Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M	10.4 GB	Fits	~107–167 tok/sestimate
Gemma 4 12B IT11.95B · Q4_K_M	8.2 GBcontext memory estimated	Fits	~123–194 tok/sestimate
Qwen3.5 9B9B · Q4_K_M	7.0 GBcontext memory estimated	Fits	~155–243 tok/sestimate
Granite 4.2 8B8.8B · Q4_K_M	7.9 GB	Fits	~159–249 tok/sestimate
Gemma 4 E4B IT8B · Q4_K_M	6.5 GBcontext memory not counted	Fits	~289–454 tok/sestimate
Qwen3 8B8.2B · Q4_K_M	7.3 GB	Fits	200 tok/smedian of 1 run
Llama 3.1 8B Instruct8B · Q4_K_M	7.1 GB	Fits	200 tok/smedian of 3 runs
Qwen3.5 4B4B · Q4_K_M	4.3 GBcontext memory estimated	Fits	~321–503 tok/sestimate

Measured speeds

15 published runs with one user, plain decoding and at most 4,096 tokens already in context, on RTX 5090. Runs on RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload) with the same GPUs count here too, and say so.

Plain single-user runs
Model	Quant	Backend	Generation	Prompt	Context	Measured on	Source
Gemma 4 26B-A4B IT	Q4_K	llama.cpp (build not stated)	180 tok/s	8,799 tok/s4,096-token prompt	4,096	RTX 5090	review2026-03
Gemma 4 26B-A4B IT	INT4 experts / HQQ4 attention / k4v4 KV	Krasis (MoE runtime; version not stated in table)	64 tok/s	5,599 tok/s	0	RTX 5090	independent benchmark2026-10
Gemma 4 31B IT	Q4_K	llama.cpp (build not stated)	61 tok/s	3,395 tok/s4,096-token prompt	4,096	RTX 5090	review2026-03
gpt-oss-20b	MXFP4 MoE	llama.cpp b6782 CUDA,BLAS	299 tok/s	13,002 tok/s2,048-token prompt	0	RTX 5090	llama.cpp benchmark discussion2025-10
gpt-oss-20b	Q4_K	llama.cpp (build not stated)	298 tok/s	9,444 tok/s4,096-token prompt	4,096	RTX 5090	review2026-03
gpt-oss-20b	MXFP4 MoE	llama.cpp b6210 CUDA	283 tok/s	9,848 tok/s2,048-token prompt	0	RTX 5090	llama.cpp benchmark discussion2025-08
gpt-oss-20b	MXFP4	Ollama (version not stated)	205 tok/s	8,519 tok/s	0	RTX 5090	review2025-10
Llama 3.1 8B Instruct	Q4_K_M	llama.cpp b11741	278 tok/s	–	0	RTX 5090	review2026-10
Llama 3.1 8B Instruct	Q4_K_M	Ollama (version not stated)	200 tok/s	30,982 tok/s	0	RTX 5090	review2025-10
Llama 3.1 8B Instruct	Q4_K_M	LocalScore (llamafile; version not stated on leaderboard)	66 tok/s	6,297 tok/s	0	RTX 5090	LocalScore2026-10
Llama 3.1 8B Instruct	Q8_0	Ollama (version not stated)	145 tok/s	31,442 tok/s	0	RTX 5090	review2025-10
Llama 3.1 8B Instruct	FP8	SGLang (lmsysorg/sglang:spark image)	138 tok/s	21,956 tok/s2,048-token prompt	0	RTX 5090	review2025-10
Llama 3.3 70B Instruct	IQ3_XXS	llama.cpp b11741	53 tok/s	–	0	RTX 5090	review2026-10
Qwen3 8B	Q4_K	llama.cpp b8189	200 tok/s	11,933 tok/s4,096-token prompt	4,096	RTX 5090	review2026-03
Qwen3.6 35B-A3B	INT4 experts / HQQ4 attention / k4v4 KV	Krasis (MoE runtime; version not stated in table)	117 tok/s	10,670 tok/s	0	RTX 5090	independent benchmark2026-10

With speculative decoding

A small draft model or extra prediction heads guess several tokens ahead. Generation is faster on text the guesses get right, so these are kept out of the medians above.

Runs with speculative decoding
Model	Quant	Backend	Generation	Prompt	Context	Measured on	Source
Qwen3.8 27B	NVFP4 (native, NVFP4 KV)	vLLM 0.27.1 + heavy local patches	101 tok/s	–	0	RTX 5090	independent benchmark2026-08

Deep into a long context

Measured with more than 4,096 tokens already in context, where every new token also reads the context memory.

Runs at long context
Model	Quant	Backend	Generation	Prompt	Context	Measured on	Source
Gemma 4 26B-A4B IT	Q4_K	llama.cpp (build not stated)	167 tok/s	7,733 tok/s16,384-token prompt	16,384	RTX 5090	review2026-03
Gemma 4 26B-A4B IT	Q4_K	llama.cpp (build not stated)	159 tok/s	6,293 tok/s32,768-token prompt	32,768	RTX 5090	review2026-03
Gemma 4 31B IT	Q4_K	llama.cpp (build not stated)	59 tok/s	2,794 tok/s16,384-token prompt	16,384	RTX 5090	review2026-03
Gemma 4 31B IT	Q4_K	llama.cpp (build not stated)	55 tok/s	2,229 tok/s32,768-token prompt	32,768	RTX 5090	review2026-03
gpt-oss-20b	Q4_K	llama.cpp (build not stated)	249 tok/s	7,168 tok/s16,384-token prompt	16,384	RTX 5090	review2026-03
gpt-oss-20b	Q4_K	llama.cpp (build not stated)	215 tok/s	5,183 tok/s32,768-token prompt	32,768	RTX 5090	review2026-03
Qwen3 8B	Q4_K	llama.cpp b8189	162 tok/s	8,538 tok/s16,384-token prompt	16,384	RTX 5090	review2026-03
Qwen3 8B	Q4_K	llama.cpp b8189	130 tok/s	6,034 tok/s32,768-token prompt	32,768	RTX 5090	review2026-03

Specifications

GPUs: 1 × NVIDIA GeForce RTX 5090
VRAM: 32 GB
System RAM: 64 GB
RAM bandwidth: 2 channels x 6000 MT/s x 8 bytes = 96.0 GB/s (dual-channel AM5)
Power supply: 1000 W
Estimated load: RTX 5090 575 W + Ryzen 9 9950X 170 W + 100 W = 845 W
Software: CUDA, Vulkan

Before you build it

DDR5 prices are at shortage highs: Corsair lists a 64 GB (2x32) DDR5-6000 kit at $1,274.99, so RAM can cost more than a used GPU.
NVIDIA specifies a 1,000 W PSU for an RTX 5090 system and a single 600 W 12V-2x6 cable (or 4x 8-pin adapter).

Sources

shop-us-en.amd.com: price of AMD Ryzen 9 9950X (16-core, 170 W)
eshop.asus.com: price of ASUS TUF GAMING X870E-PLUS WIFI7
corsair.com: price of Corsair Vengeance 64GB (2x32GB) DDR5-6000 CL30
corsair.com: price of Corsair RM1000x (1000 W)
corsair.com: price of Corsair 7000D AIRFLOW full tower (450 mm GPU clearance, 8 slots)
corsair.com: price of Corsair MP600 CORE XT 2TB PCIe 4.0 NVMe
nvidia.com: RTX 5090 575 W, 1000 W system PSU

Compared with

Similar machines

RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload)~$3,763 + unpriced parts
1x RTX 3090 (used)~$2,944 + unpriced parts
1x RTX PRO 6000 Blackwell~$20,833