GeForce RTX 5090
32 GB GDDR7 · 1,792 GB/s · 575 W
RTX 5090 32 GB · 64 GB RAM → Mixtral 8x7B Instruct
GPU · NVIDIA
RTX 5090 for local LLMs: which models run?
With 32 GB of VRAM and 1792 GB/s of bandwidth, the RTX 5090 runs models up to about Mixtral 8x7B Instruct entirely on the GPU.
- VRAM32 GB GDDR7
- Bandwidth1,792 GB/s
- Board power575 W
- ArchitectureBlackwell
Specifications
- VRAM
- 32 GB
- Memory type
- GDDR7
- Memory bus
- 512-bit
- Bandwidth
- 1,792 GB/s
- Board power
- 575 W
- Backends
- CUDA, Vulkan
- Interface
- PCIe 5.0 x16
- Released
- January 2025
- Segment
- Consumer
- FP16 compute
- Not available
Official specData updated 24 days agoSource: NVIDIA specifications
What the RTX 5090 runs
| Qwen3 30B-A3BQ6_K · 8K~26.1 GB · 142–250 tok/s estimate | Excellent |
|---|---|
| Qwen3 Coder 30B-A3B InstructQ6_K · 8K~26.1 GB · 142–250 tok/s estimate | Excellent |
| Gemma 3 27B ITQ6_K · 8K~23.8 GB · 50–71 tok/s estimate | Excellent |
| Gemma 2 27B ITQ6_K · 8K~25.2 GB · 51–70 tok/s estimate | Excellent |
| gpt-oss-20bMXFP4 · 8K~13 GB · 264–291 tok/s measured | Excellent |
| Qwen3 14BQ8_0 · 8K~17.5 GB · 70–100 tok/s estimate | Excellent |
| DeepSeek R1 Distill Qwen 14BQ8_0 · 8K~17.8 GB · 68–97 tok/s estimate | Excellent |
| Qwen2.5 14B InstructQ8_0 · 8K~17.8 GB · 68–97 tok/s estimate | Excellent |
| Phi-4 14BQ8_0 · 8K~17.6 GB · 70–99 tok/s estimate | Excellent |
| Mistral Nemo 12B InstructQ8_0 · 8K~14.9 GB · 81–117 tok/s estimate | Excellent |
| Gemma 3 12B ITQ8_0 · 8K~14.1 GB · 82–121 tok/s estimate | Excellent |
| Gemma 2 9B ITQ8_0 · 8K~12.7 GB · 96–143 tok/s estimate | Excellent |
| Qwen3 8BQ8_0 · 8K~10.5 GB · 113–167 tok/s estimate | Excellent |
| Llama 3.1 8B InstructQ8_0 · 8K~10.2 GB · 118–173 tok/s estimate | Excellent |
| DeepSeek R1 Distill Llama 8BQ8_0 · 8K~10.2 GB · 118–173 tok/s estimate | Excellent |
| Qwen2.5 7B InstructQ8_0 · 8K~9.2 GB · 130–190 tok/s estimate | Excellent |
| Mistral 7B Instruct v0.3Q8_0 · 8K~9.2 GB · 127–189 tok/s estimate | Excellent |
| Gemma 3 4B ITQ8_0 · 8K~5.4 GB · 195–322 tok/s estimate | Excellent |
| Qwen3 4BQ8_0 · 8K~6.2 GB · 176–287 tok/s estimate | Excellent |
| Phi-4 Mini Instruct 3.8BQ8_0 · 8K~6 GB · 190–308 tok/s estimate | Excellent |
| Llama 3.2 3B InstructQ8_0 · 8K~5.1 GB · 223–363 tok/s estimate | Excellent |
| Qwen3 1.7BQ8_0 · 8K~3.7 GB · 299–537 tok/s estimate | Excellent |
| Llama 3.2 1B InstructQ8_0 · 8K~2.5 GB · 497–876 tok/s estimate | Excellent |
| Qwen3 0.6BQ8_0 · 8K~2.5 GB · 403–843 tok/s estimate | Excellent |
| Mixtral 8x7B InstructQ4_K_M · 8K~27.5 GB · 103–154 tok/s estimate | Compatible |
| Qwen3 32BQ6_K · 8K~29.2 GB · 42–59 tok/s estimate | Compatible |
| DeepSeek R1 Distill Qwen 32BQ6_K · 8K~29.3 GB · 42–59 tok/s estimate | Compatible |
| Qwen2.5 32B InstructQ6_K · 8K~29.3 GB · 42–59 tok/s estimate | Compatible |
| Qwen2.5 Coder 32B InstructQ6_K · 8K~29.3 GB · 42–59 tok/s estimate | Compatible |
| QwQ 32BQ6_K · 8K~29.3 GB · 42–59 tok/s estimate | Compatible |
| Mistral Small 3.2 24B InstructQ8_0 · 8K~26.6 GB · 48–65 tok/s estimate | Compatible |
| Llama 4 Scout 17B-16E InstructQ3_K_M · 8K~55.7 GB · 4.5–9 tok/s estimate | CPU/GPU offload |
| Qwen2.5 72B InstructQ3_K_M · 8K~38.2 GB · 2.5–5 tok/s estimate | CPU/GPU offload |
| Llama 3.3 70B InstructQ3_K_M · 8K~37 GB · 3–6 tok/s estimate | CPU/GPU offload |
| DeepSeek R1 Distill Llama 70BQ3_K_M · 8K~37 GB · 3–6 tok/s estimate | CPU/GPU offload |
| gpt-oss-120bMXFP4 · 8K~63.2 GB · 13–24 tok/s estimate | Possible but slow |
| DeepSeek R1 (671B)Q8_0 · 8K~699 GB · No data | Incompatible |
| Qwen3 235B-A22BQ8_0 · 8K~247 GB · No data | Incompatible |
For each model, the highest-precision quantization that runs entirely in fast memory; otherwise the one with the best result. Speeds are ranges; "estimate" unless a benchmark exists for this exact hardware.
EstimateData updated 24 days agoSource: Run LLM Locally engine
Limits in plain words
32 GB of VRAM holds the weights and KV cache of models up to about Mixtral 8x7B Instruct at Q4_K_M. Bigger models spill into system RAM and slow down sharply: generation speed is then limited by RAM bandwidth, not by the GPU.
Where to buy a RTX 5090
Prices and links follow the market you choose.
No current offer imported for this market: the link opens a marketplace search.
Frequently asked questions
Which LLMs can the RTX 5090 run?
Entirely on the GPU: Mixtral 8x7B Instruct (Q4_K_M), Qwen3 32B (Q6_K), DeepSeek R1 Distill Qwen 32B (Q6_K), Qwen2.5 32B Instruct (Q6_K), Qwen2.5 Coder 32B Instruct (Q6_K), QwQ 32B (Q6_K), Qwen3 30B-A3B (Q6_K), Qwen3 Coder 30B-A3B Instruct (Q6_K). Larger models need CPU/GPU offload.
Can the RTX 5090 run a 70B model?
Possible but slow: a 70B model at Q4_K_M needs about 45.1 GB; the RTX 5090 has 32 GB.
How is the speed estimated?
Token generation reads every active weight once per token, so it is limited by memory bandwidth (1792 GB/s here). We apply an efficiency range per backend and label the result as an estimate until real benchmarks exist.