What LLMs Can the RTX 5090 32GB Run?

The RTX 5090 32GB has 32 GB of VRAM and 1792 GB/s of bandwidth. It runs 42 of 58 models fully at Q4_K_M with 8K context, up to Mixtral 8x7B, and generates about 202 tokens/s on an 8B model.

NVIDIA · Blackwell · Released 2025-01 · Updated October 14, 2026

42 of 58 models fit

RTX 5090 32GB (32.0 GB) at Q4_K_M with 8K context. Largest that fits fully: Mixtral 8x7B.

ModelNeedsStatusSpeedBest quant that fits
Mixtral 8x7B28.0 GBFits~96 tok/sQ4_K_M
Qwen 2.5 32B21.4 GBFits~54 tok/sQ6_K
Qwen 2.5 Coder 32B21.4 GBFits~54 tok/sQ6_K
DeepSeek R1 Distill Qwen 32B21.4 GBFits~54 tok/sQ6_K
Qwen 3 32B21.4 GBFits~54 tok/sQ6_K
Qwen 3.5 35B-A3B21.0 GBFits~176 tok/sQ6_K
Qwen 3.6 35B-A3B21.0 GBFits~176 tok/sQ6_K
Gemma 4 31B19.8 GBFits~57 tok/sQ6_K
Qwen 3 30B-A3B18.8 GBFits~146 tok/sQ6_K
Qwen3-Coder 30B-A3B18.8 GBFits~146 tok/sQ6_K
Gemma 3 27B17.6 GBFits~64 tok/sQ8_0
Qwen 3.5 27B16.9 GBFits~67 tok/sQ8_0
Qwen 3.6 27B16.9 GBFits~67 tok/sQ8_0
Qwen 3.8 27B16.9 GBFits~67 tok/sQ8_0
Mistral Small 3.1 24B15.6 GBFits~74 tok/sQ8_0
Gemma 4 26B-A4B15.6 GBFits~191 tok/sQ8_0
gpt-oss-20b12.9 GBFits~229 tok/sMXFP4 (official)
Qwen 2.5 14B10.7 GBFits~113 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 14B10.7 GBFits~113 tok/sFP16 / BF16
Phi-4 14B10.6 GBFits~114 tok/sFP16 / BF16
Qwen 3 14B10.5 GBFits~115 tok/sFP16 / BF16
Ministral 3 14B9.7 GBFits~125 tok/sFP16 / BF16
Mistral Nemo 12B9.0 GBFits~137 tok/sFP16 / BF16
Gemma 3 12B8.8 GBFits~136 tok/sFP16 / BF16
Ministral 3 8B6.7 GBFits~191 tok/sFP16 / BF16
Qwen 3 8B6.6 GBFits~195 tok/sFP16 / BF16
Qwen 3.5 9B6.4 GBFits~195 tok/sFP16 / BF16
Llama 3.1 8B6.3 GBFits~202 tok/sFP16 / BF16
DeepSeek R1 Distill Llama 8B6.3 GBFits~202 tok/sFP16 / BF16
Mistral 7B v0.35.7 GBFits~220 tok/sFP16 / BF16
Gemma 4 E4B5.7 GBFits~215 tok/sFP16 / BF16
Qwen 2.5 7B5.6 GBFits~225 tok/sFP16 / BF16
DeepSeek R1 Distill Qwen 7B5.6 GBFits~225 tok/sFP16 / BF16
Qwen 3 4B4.2 GBFits~337 tok/sFP16 / BF16
Phi-4 Mini4.1 GBFits~360 tok/sFP16 / BF16
Gemma 4 E2B4.0 GBFits~330 tok/sFP16 / BF16
Gemma 3 4B3.7 GBFits~363 tok/sFP16 / BF16
Qwen 3.5 4B3.7 GBFits~375 tok/sFP16 / BF16
Ministral 3 3B3.5 GBFits~413 tok/sFP16 / BF16
Llama 3.2 3B3.5 GBFits~425 tok/sFP16 / BF16
Llama 3.2 1B1.7 GBFits~1086 tok/sFP16 / BF16
Gemma 3 1B1.6 GBFits~1243 tok/sFP16 / BF16
Qwen 3.5 122B-A10B70.2 GB43% on GPU~11 tok/s–
Mistral Small 4 119B68.2 GB44% on GPU~17 tok/s–
gpt-oss-120b67.2 GB45% on GPU~21 tok/s–
Llama 4 Scout64.0 GB47% on GPU~6.7 tok/s–
GLM-4.5-Air62.1 GB48% on GPU~9.4 tok/s–
Qwen3-Next 80B-A3B46.3 GB65% on GPU~44 tok/sQ2_K
Qwen 2.5 72B44.5 GB68% on GPU~2.7 tok/sQ2_K
Llama 3.3 70B43.2 GB70% on GPU~2.9 tok/sQ2_K
DeepSeek R1 Distill Llama 70B43.2 GB70% on GPU~2.9 tok/sQ2_K
Kimi K2.6566 GBToo small––
DeepSeek R1 671B380 GBToo small––
Llama 3.1 405B234 GBToo small––
Qwen 3.5 397B-A17B226 GBToo small––
DeepSeek V4 Flash161 GBToo small––
Qwen 3 235B-A22B135 GBToo small––
MiniMax-M2133 GBToo small––

RTX 5090 32GB for local AI

The first consumer card with 32 GB and nearly double the 4090's bandwidth. It runs 32B models at Q6_K or with long context, and 70B models only at very low quantization.

Setup tips: plan for a quality 1,000 W+ power supply and good case airflow; a power limit of 450–500 W keeps LLM speed nearly unchanged. 32 GB fits 27–32B models at Q6_K or with 64K+ contexts, and makes FLUX.2 practical with GGUF weights.

Among the 2 devices we track with about 32 GB of usable memory, the RTX 5090 32GB is the fastest at generating text, because generation speed follows memory bandwidth. Others in the same class: Radeon AI PRO R9700 32GB. They all hold the same models; the faster ones simply answer sooner.

Our picks for the RTX 5090 32GB

The model we would run first for each job, chosen from our recommendations by what fits fully with an 8K context. "Best quant" is the highest quality that fits; "max context" is at Q4_K_M or the native format.

TaskModelBest quantMax contextSpeed
Best all-rounderQwen 3.6 27BQ8_0219K~67 tok/s
CodingQwen 3.6 27BQ8_0219K~67 tok/s
Fastest repliesQwen 3.6 35B-A3BQ6_K256K~176 tok/s
Long documents (64K+)Qwen 3.8 27BQ8_0219K~67 tok/s
Images in, text outGemma 4 31BQ6_K139K~57 tok/s
Image generationFLUX.2 [dev]GGUF Q6_K––
SpecValue
Memory32 GB
Memory bandwidth1792 GB/s
ArchitectureBlackwell
Released2025-01
Llama 3.1 8B Q4_K_M speed~202 tok/s
Largest model that fitsMixtral 8x7B

Image and video models on the RTX 5090 32GB

Peak VRAM with model swapping (ComfyUI default on smaller cards).

ModelFP8FP8 fitGGUF Q4Q4 fit
Stable Diffusion 1.51.8 GBFits1.7 GBFits
SDXL 1.04.1 GBFits4.1 GBFits
Stable Diffusion 3.5 Medium6.4 GBFits6.4 GBFits
Stable Diffusion 3.5 Large8.6 GBFits6.4 GBFits
FLUX.1 [dev]12.1 GBFits7.3 GBFits
FLUX.1 [schnell]12.1 GBFits7.3 GBFits
Qwen-Image20.4 GBFits12.3 GBFits
Wan 2.2 A14B (T2V / I2V)18.8 GBFits13.2 GBFits
Wan 2.2 TI2V 5B8.0 GBFits6.5 GBFits
HunyuanVideo18.3 GBFits13.1 GBFits
FLUX.1 Kontext [dev]12.3 GBFits7.5 GBFits
FLUX.2 [dev]31.0 GBTight fit23.6 GBFits
Z-Image Turbo6.8 GBFits4.9 GBFits
Qwen-Image-Edit 250920.5 GBFits12.4 GBFits
LTX-219.2 GBFits12.6 GBFits

Upgrade path

The next step up is the RTX A6000 48GB with 48 GB.

Comparisons

RTX 5090 32GB for LLMs: common questions

Can the RTX 5090 32GB run Llama 3.3 70B?

Only partly. It needs about 43.2 GB at Q4_K_M, so about 70% fits in VRAM and the rest runs from system RAM, which is slow. Q2_K fits fully.

Can the RTX 5090 32GB run Qwen 3 32B?

Yes. Qwen 3 32B needs about 21.4 GB at Q4_K_M with 8K context, and the best quality that fits is Q6_K.

Can the RTX 5090 32GB run Qwen 3 14B?

Yes. Qwen 3 14B needs about 10.5 GB at Q4_K_M with 8K context, and the best quality that fits is FP16 / BF16.

How fast is the RTX 5090 32GB for LLMs?

About 202 tokens per second for an 8B model at Q4_K_M, based on its 1792 GB/s memory bandwidth. Larger models scale down roughly in proportion to their size.

Related