llmrun

NVIDIA · Blackwell

Best AI Models for NVIDIA GeForce RTX 5090 (32 GB)

Runs 27–31B and 35B MoE models well; 70B doesn't fit in 32 GB.

32 GB GDDR71792 GB/s210 TFLOPS FP16575 WBlackwell 2025
Buyabout $5,376

What LLMs can the RTX 5090 run?

#ModelSpeedllmrun ScoreQuant · VRAM
1
Qwen3.8 27BAlibabaHow to run →
Q4_K_M · 17.9 GB · 66.9 t/s
66.9 tok/sQ4_K_M · 17.9 GB
2
Qwen3.6 27BAlibaba
Q4_K_M · 17.9 GB · 66.9 t/s
66.9 tok/sQ4_K_M · 17.9 GB
3
Gemma 4 31B ITGoogle
Q4_K_M · 21.7 GB · 57.1 t/s
57.1 tok/sQ4_K_M · 21.7 GB
4
Gemma 4 26B A4B ITGoogle
Q4_K_M · 16.5 GB · 250 t/s
250 tok/sQ4_K_M · 16.5 GB
5
NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA
Q4_K_M · 19.3 GB · 244 t/s
244 tok/sQ4_K_M · 19.3 GB
See all 346 models on the RTX 5090 →Best quant that fits · speeds estimated on this card

About the RTX 5090

The RTX 5090 is the quickest card in the 24–32 GB class: it runs 27–31B and 35B MoE models well. It is the most efficient per watt in its tier, and pricier per token than most of the cards compared at ≈$5,376. Blackwell generation, released 2025.

Speed in 24–32 GB#11792 GB/s
Per watt#10.32 tok/s per W on Qwen3.5 9B (9B params)
Per dollar#73.4 tok/s per $100 on Qwen3.5 9B (9B params)
GenerationBlackwellreleased 2025

RTX 5090 strengths and weaknesses

Strengths

  • Fastest card here: 78% faster than the RTX 4090, 183 vs 103 tok/s on Qwen3.5 9B
  • Runs 35B MoE models: Qwen3.6 35B A3B at 213 tok/s in 22 GB
  • Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 14.1 GB free
  • Reads long prompts fast: A 4K-token prompt on Qwen3.8 27B starts answering in 2.4 s

Weaknesses

  • 70B models don't fit: Llama 3.3 70B Instruct needs 51.3 GB at Q4 vs 32 GB here
  • 575 W under load: Plan a 1,150 W PSU; ≈$126 a year at 4 h a day
  • Priced well above MSRP: ≈$5,376 new vs its $1,999 MSRP (+169%)

Speed per dollar and per watt

On Qwen3.5 9B (9B params) at Q4_K_M. New prices are compared with new, used with used. The last two columns are the speed you get per $100 of price and per 100 W of board power; higher is better.

CardNew pricetok/stok/s per $100tok/s per 100 W
RX 7900 XTX24 GB
about $1,13090.6
8.0
25.5
RTX 5070 Ti16 GBBest new value
about $1,17091.6
7.8
30.5
RX 907016 GB
about $80060.4
7.6
27.5
RX 9070 XT16 GB
about $81560.4
7.4
19.9
RTX 508016 GB
about $1,69098.1
5.8
27.3
RX 9060 XT 16GB16 GB
about $59030.2
5.1
18.9
RTX 509032 GB
about $5,376183
3.4
31.8
CardUsed pricetok/stok/s per $100tok/s per 100 W
RX 907016 GBBest used value
about $65060.4
9.3
27.5
RX 7900 XTX24 GB
about $1,00090.6
9.1
25.5
RTX 5070 Ti16 GB
about $1,10091.6
8.3
30.5
RTX 3090 Ti24 GB
about $1,600103
6.4
22.9
RTX 309024 GB
about $1,50095.7
6.4
27.3
RTX 508016 GB
about $1,64098.1
6.0
27.3
RTX 409024 GB
about $3,000103
3.4
22.9
RTX 509032 GB
about $6,400183
2.9
31.8

Cost per token: the RTX 5090 against 16–32 GB cards

On Qwen3.5 9B (9B params, Q4_K_M, 4K context). Speed, cost per token and power per token depend on the model.

16 GB24 GB32 GBthis card
$0.00$0.50$1.00$1.50$2.00$2.50$0$2,000$4,000$6,000$ per 1M output tokens · lower is betternew price →RX 7900 XTX · 24 GB · ≈$1,130 · $0.95 per 1M tokensRTX 5070 Ti · 16 GB · ≈$1,170 · $0.95 per 1M tokensRX 9070 · 16 GB · ≈$800 · $0.99 per 1M tokensRTX 5080 · 16 GB · ≈$1,690 · $1.25 per 1M tokensRX 9070 XT · 16 GB · ≈$815 · $1.07 per 1M tokensRX 9060 XT 16GB · 16 GB · ≈$590 · $1.46 per 1M tokensRTX 5090 · 32 GB · ≈$5,376 · $1.99 per 1M tokensRTX 5090RTX 5070 TiRX 9070RTX 5080RX 9070 XTRX 9060 XT 16GB
$0.00$0.50$1.00$1.50$2.00$2.50$0$2k$4k$6k$ per 1M output tokens · lower is betternew price →RX 7900 XTX · 24 GB · ≈$1,130 · $0.95 per 1M tokensRTX 5070 Ti · 16 GB · ≈$1,170 · $0.95 per 1M tokensRX 9070 · 16 GB · ≈$800 · $0.99 per 1M tokensRTX 5080 · 16 GB · ≈$1,690 · $1.25 per 1M tokensRX 9070 XT · 16 GB · ≈$815 · $1.07 per 1M tokensRX 9060 XT 16GB · 16 GB · ≈$590 · $1.46 per 1M tokensRTX 5090 · 32 GB · ≈$5,376 · $1.99 per 1M tokensRTX 5090RTX 5070 TiRTX 5080
Cost per 1M output tokenson Qwen3.5 9B · card over 3 years + power
$1.99
Card, new price
$1.86
Power
$0.13
3-year cost of ownership$5,754
Or rent in the cloud$0.50/hr

Renting costs as much as buying after ≈10,800 hours (≈7.4 years at 4 h a day).

Include card price

Each dot is a card; hover or tap it for its numbers, click to open it. Cards without a price for the chosen basis are left out. How we calculate cost per token

Similar GPUs

Cards people weigh against the RTX 5090, each with the one reason it is worth a look. Speeds on Qwen3.5 9B.

Slower but $4,246 cheaperRX 7900 XTXAMD · 24 GB · RDNA 3
▼ 50% speed▼ $4,246 less▼ 8 GBabout $1,130
44% slower, 24 GBRTX 409024 GB
▼ 44% speed▼ 8 GBabout $1,599rent $0.34/hr
44% slower, 24 GBRTX 3090 Ti24 GB
▼ 44% speed▼ 8 GBabout $1,999rent $0.27/hr

RTX 5090 tokens per second by model size

Output speed depends on memory bandwidth (1792 GB/s). Grey rows need more than 32 GB.

  1. 9BQwen3.5 9B · Q4_K_M · 6.6 GB
    183tok/s
  2. 14BQwen3 14B · Q4_K_M · 11.9 GB
    123tok/s
  3. 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
    66.9tok/s
  4. 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
    213tok/s
  5. 70BLlama 3.3 70B Instruct · not at 4-bit
    needs 51.3 GB
  6. 120B MoEGPT OSS 120B · not at 4-bit
    needs 70.8 GB

Estimated speeds: Q4_K_M, 4K context, one user. How we estimate

All 1176 models on the RTX 5090

Featured only

Best quant that fits, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works

LLM models compatible with NVIDIA GeForce RTX 5090 — ranked by Score, recency and Fit
ModelVRAMScoreFit
SmolLM3 3B Base
Q4_K_M·506.4 t/s
3.4 GBS100, Runs great
Qwen3 4B Thinking 2507
Q4_K_M·401.7 t/s
4.2 GBS100, Runs great
MiniCPM MoE 8x2B
Q4_K_M·170.0 t/s
11.4 GBS98, Runs great
Gpt2 Xl
Q4_K_M·1098.9 t/s
1.1 GBS100, Runs great
Pythia 1.4B
Q4_K_M·1164.8 t/s
1 GBS100, Runs great
TinyLLama V0
Q4_K_M·116480.0 t/s
0.0 GBS100, Runs great
Nemotron Orchestrator 8B
Q4_K_M·211.0 t/s
7.6 GBS100, Runs great
EXAONE 3.5 2.4B Instruct
Q4_K_M·732.6 t/s
3.4 GBS100, Runs great
NuExtract
Q4_K_M·342.6 t/s
4.2 GBS100, Runs great
Olmo 3 1025 7B
Q4_K_M·202.6 t/s
8.4 GBS100, Runs great
GPT J 6B
Q4_K_M·294.1 t/s
4.0 GBS100, Runs great
GOT OCR2 0
Q4_K_M·1252.5 t/s
2.3 GBS100, Runs great
Llama 68M
Q4_K_M·29120.0 t/s
0.0 GBS100, Runs great
Vicuna 33B V1.3
Q4_K_M·53.5 t/s
21.8 GBA81, Runs well
Llama 3.1 8B
Q4_K_M·219.8 t/s
7.9 GBS100, Runs great

RTX 5090 specs, power and PSU

Specs

Architecture
Blackwell
VRAM
32 GB GDDR7
Memory bandwidth
1792.0 GB/s
FP16 (dense)
209.50 TFLOPS
CUDA cores
21,760
Tensor cores
680
Compute capability
10.0
TDP
575 W
Released
2025-01-30
MSRP
$1,999

RTX 5090 FAQ

Is the RTX 5090 still worth it in 2026?

At ≈$5,376 you get 3.4 tok/s per $100 on Qwen3.5 9B, which ranks 7th of the 7 cards compared here; it is also the fastest out of 6 cards with 24–32 GB. It runs 27–31B and 35B MoE models well; 70B doesn't fit in 32 GB.

What is the largest model the RTX 5090 can run?

Qwen3.6 35B A3B (36.0B) is the largest at 4-bit or better: at Q4_K_M it uses 22.0 GB of 32 GB, and runs at about 213 tok/s.

Can the RTX 5090 run Qwen3.8 27B?

Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of 32 GB, and it is the best model that runs well on this card.

Can the RTX 5090 run 70B models?

No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 32 GB on this card.

How many tokens per second does the RTX 5090 give?

About 66.9 tok/s on Qwen3.8 27B at Q4_K_M, with the first token after 2.4 s on a 4K-token prompt.

What power supply does the RTX 5090 need?

Plan a 1150 W PSU (2× the 575 W TDP).

Is 32 GB enough for coding and agents?

Yes for coding: Qwen3.8 27B, the best of the 10 coding models that fit, runs at about 67 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.

Is renting cheaper than buying an RTX 5090?

Rent from $0.50/hr. Renting costs as much as buying after ≈10,800 hours (≈7.4 years at 4 h a day).

Related

Best coding LLMs24 GB VRAM48 GB VRAMGPUs with 32 GB or more