NVIDIA
Operating mode
Use this to bias workload recommendations toward responsiveness, background autonomy, lighter serving, or multi-GPU scale-out.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
The RTX 5090 is NVIDIA's flagship consumer GPU built on the Blackwell architecture. With 32 GB of next-generation GDDR7 memory and 21,760 CUDA cores, it represents a generational leap in local AI capability. Its 1,792 GB/s memory bandwidth enables exceptionally fast token generation, and the 32 GB VRAM pool can handle 70B+ parameter models with comfortable quantization headroom.
Official product page ↗Beyond LLMs
What AI tasks this GPU can handle — from text generation to image and video creation.
| Capability | Status | Representative Model |
|---|---|---|
| LLM Chat (7B) | Runs natively | Llama 3.1 8B Q4 |
| LLM Coding (30B) | Runs natively | Qwen 3 30B Q4 |
| LLM Large (70B) | Won’t fit | Llama 3.1 70B Q4 |
| Image Gen (SDXL) | Runs natively | SDXL 1.0 FP16 |
| Image Gen (Flux) | Runs natively | Flux.1 Dev FP16 |
| Image Gen (SD 3.5) | Runs natively | SD 3.5 Large FP16 |
| Video Short (25f) | Runs natively | LTX Video 2B |
| Video Long (100f) | Won't fit | Wan Video 14B |
Architecture
Blackwell is NVIDIA's fifth-generation RTX architecture, built on TSMC's 4NP process. It introduces 5th-generation Tensor Cores with native FP4 precision support, enabling double the inference throughput per watt compared to Ada Lovelace's FP8 operations. Key innovations include the Neural Rendering Pipeline for AI-driven shading and the debut of GDDR7 memory in consumer GPUs.
AI Relevance
FP4 Tensor Cores deliver the highest tokens-per-watt efficiency in any consumer architecture. Native FP4 quantization means models can run at lower precision with minimal quality loss, effectively doubling the effective VRAM for model weights.
Blackwell is NVIDIA's fifth-generation RTX architecture, built on TSMC's 4NP process. It introduces 5th-generation Tensor Cores with native FP4 precision support, enabling double the inference throughput per watt compared to Ada Lovelace's FP8 operations.
The RTX 5090 uses the full GB202 GPU die with 170 Streaming Multiprocessors housing 21,760 CUDA cores and 680 Tensor Cores. The new Neural Rendering Pipeline integrates AI-driven shading and material synthesis directly into the graphics pipeline.
The memory subsystem marks the debut of GDDR7 in consumer GPUs. Running at 28 Gbps on a 512-bit bus, it delivers 1,792 GB/s of bandwidth — a 78% improvement over the RTX 4090. For LLM inference, this translates directly to faster autoregressive decoding since token generation is memory-bandwidth-bound.
Kaufberatung
Ausgezeichnete Wahl für lokale KI
Führt 26 von 50 Top-Modellen gut aus — ein starker Allrounder für lokale Inferenz.
32.0 GB
VRAM
$1,999
UVP
$62/GB
Kosten pro GB VRAM
Beste Modelle für diese GPU
What will limit you first
This setup is broadly balanced for this model.
No major red flags
This recommendation has enough memory headroom and acceptable estimated speed for the selected workload.
Best upgrade itinerary
Unlocks 11 additional models that do not fit on the current setup.
Mehr Spielraum gewünscht? MacBook Pro M1 Max 64GB (64.0 GB unified memory) ist die nächste Stufe.
Cost vs cloud API
Assumes 4 hours/day of active inference at 131 tok/s, RTX 5090 32GB amortized over 36 months, US residential electricity ($0.15/kWh), blended cloud pricing at $10 per 1M tokens (GPT-4o / Claude Sonnet tier).
56.5M
Tokens/month at this pace
$63.6
Monthly local cost
$565
Same tokens on cloud API
$1.13
Local $/1M tokens
Break-even: pays for itself in 3.6 months vs cloud API at this workload. Price reference: $2.0k MSRP.
Chat
SQwen 3.5 35B A3B matches Chat and keeps a practical fit profile. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, ollama, lm-studio.
Coding
SQwen 3.6 27B is a specialized fit for Coding. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
Agentic Coding
SQwen 3.6 27B is a specialized fit for Agentic Coding. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
Reasoning
SDevstral Small 2 24B Instruct matches Reasoning and keeps a practical fit profile. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, ollama, lm-studio.
RAG
SQwen 3.5 27B matches RAG and keeps a practical fit profile. It is a recent-generation family, which helps on current local SOTA workloads. It should run, but memory headroom will be limited. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, ollama, lm-studio.
Qwen3-Coder 30B A3B InstructQwen3-VL 30B A3B InstructQwen 3.5 27BQwen 3 30B A3B
Magistral Small 2507Devstral Small 2 24B InstructQwen 3.6 35B A3BQwen 3.6 27BQwen 3.5 35B A3BNemotron 3 Nano 30BNemotron Cascade 2 30B A3BDevstral Small 1.1
GPT-OSS 20BQwen 3 14BQwen 3 32B
Phi-4-reasoning-plus 14BQwen 3.5 9BQwen 3 8BQwen 3.5 4BMinistral 3 14B
EXAONE 4.0 32BNemotron Nano 8BPhi-4 Mini Reasoning 4B
Jina Embeddings v3
BGE M3
Gemma 4 31BQwen 3.5 397B A17BDevstral 2 123B Instruct
Kimi K2.5Kimi K2.6
DeepSeek V4 ProQwen 3.5 122B A10BDeepSeek V4 FlashMistral Small 4 119B
Command A 111BQwen 2.5 VL 72BGPT-OSS 120B
GLM-5.2Qwen3-Coder-NextGLM-5.1Pixtral Large 124BGLM-5DeepSeek V3.2Qwen 3 235B A22BQwen3-Coder 480B A35B InstructLeanstral 119B A6BDeepSeek Coder V2 236BDeepSeek R1 671BDeepSeek V3.1 671BFast erreichbar
Hochwertige Modelle, die etwas mehr Speicher benötigen
Image & Video Generation
46 of 56 models can generate images or video on your RTX 5090 32GB
| Model | Max Resolution | Gen Time | Grade |
|---|---|---|---|
| SD TurboImage | 512×512 | 400ms | S |
| Stable Diffusion 1.5Image | 512×768 | 900ms | S |
| Realistic Vision v5.1Image | 512×768 | 900ms | S |
| DreamShaper 8Image | 512×768 | 900ms | S |
| LCM DreamShaper v7Image | 512×768 | 300ms | S |
| PixArt-SigmaImage | 1024×1024 | ~3.5s | S |
| FramePack I2VVideo | 256×256 | ~6.5s/frame | S |
| SDXL TurboImage | 512×512 | 400ms | S |
| SDXL LightningImage | 1024×1024 | ~1.3s | S |
| Stable Diffusion XL 1.0Image | 1024×1024 | ~3.5s | S |
| Playground v2.5Image | 1024×1024 | ~5.3s | S |
| RealVisXL v5.0Image | 1024×1024 | ~4s | S |
| DreamShaper XLImage | 1024×1024 | ~4s | S |
| Juggernaut XL v9Image | 1024×1024 | ~4s | S |
| Animagine XL 3.1Image | 1024×1024 | ~4s | S |
| Pony Diffusion V6 XLImage | 1024×1024 | ~4s | S |
| Animagine XL 4.0Image | 1024×1024 | ~4s | S |
| Illustrious XLImage | 1024×1024 | ~4s | S |
| Wan Video 2.1 1.3BVideo | 480×832 | ~2.6s/frame | S |
| Stable Diffusion 3.5 MediumImage | 1024×1024 | ~6.2s | S |
| Flux.2 Klein 4BImage | 1024×1024 | ~1.1s | S |
| LTX Video 2BVideo | 1280×720 | ~3.1s/frame | S |
| KolorsImage | 1024×1024 | ~7.1s | S |
| Stable CascadeImage | 1024×1024 | ~8.9s | S |
| AuraFlow v0.3Image | 1536×1536 | ~15.9s | S |
| Stable Diffusion 3.5 LargeImage | 1024×1024 | ~19.5s | S |
| Stable Diffusion 3.5 Large TurboImage | 1024×1024 | ~3.5s | S |
| CogVideoX 2BVideo | 720×480 | ~3.1s/frame | S |
| HunyuanVideoVideo | 256×256 | ~6.5s/frame | S |
| ChromaImage | 1024×1024 | ~3.5s | S |
| Z-Image TurboImage | 1536×1536 | ~3.7s | S |
| Flux.1 DevImage | 256×256 | ~27.9s | S |
| Flux.1 SchnellImage | 256×256 | ~5.4s | S |
| LTX Video 13BVideo | 256×256 | ~6.5s/frame | S |
| Flux.1 Kontext DevImage | 256×256 | ~31s | S |
| AnimateDiff v1.5.3Video | 512×768 | ~1.6s/frame | S |
| Cosmos Diffusion 7BVideo | 1024×576 | ~5.1s/frame | A |
| CogVideoX 5BVideo | 720×480 | ~4.4s/frame | A |
| Wan2.2 TI2V 5BVideo | 832×480 | ~4.4s/frame | A |
| Flux.2 Klein 9BImage | 1024×1024 | ~1.8s | A |
| Flux.1 Fill DevImage | 256×256 | ~26.4s | B |
| Krea 2Image | 256×256 | ~8.4s | B |
| Sulphur 2Video | 256×256 | ~10.3s/frame | B |
| Ideogram 4Image | 256×256 | ~7.9s | D |
| Mochi 1 PreviewVideo | 256×256 | ~10.5s/frame | D |
| HunyuanVideo 1.5Video | 256×256 | ~10.1s/frame | D |
| Helios 14BVideo | 256×256 | ~6.7s/frame | F |
| SkyReels V2 14BVideo | 256×256 | ~6.7s/frame | F |
| Wan Video 2.1 14BVideo | 256×256 | ~6.7s/frame | F |
| Wan Video 2.2 14BVideo | 256×256 | ~6.7s/frame | F |
| Qwen ImageImage | 256×256 | ~6s | F |
| Qwen Image EditImage | 256×256 | ~6s | F |
| LTX-2 22BVideo | 256×256 | ~8s/frame | F |
| Flux.2 DevImage | 256×256 | ~2m 48s | F |
| MAGI-1Video | 256×256 | ~8.3s/frame | F |
| HunyuanImage 3.0Image | 256×256 | ~10.5s | F |
Image models estimated at 1024×1024 (28 steps, FP16). Video models estimated at 768×512 (25 frames, 30 steps, FP16). Actual performance varies with runtime and system load.
Multi-GPU scaling
Scale out with multiple GPUs for larger models. PCIe interconnect with 28% scaling overhead.
| Config | Effective memory | Models that fit | Est. bandwidth |
|---|---|---|---|
| 1× RTX | 32 GB | 330/380 | 1,792 GB/s |
| 2× RTX | 64 GB | 348/380 | 2,580 GB/s |
Model counts use default quantization at coding workload settings. Multi-GPU scaling factor: 0.72× per additional GPU.
Upgrade paths
See what you unlock with more powerful hardware
Upgrade-Optionen
2× RTX 5090 32GBMulti-GPUUnlocks 18 additional models that do not fit on the current setup.
Scale-out only pays off if the host platform has enough PCIe lanes, slot spacing, power, and cooling.
The bigger the setup gets, the more the runtime matters. Multi-GPU and multi-user serving are where vLLM, SGLang, TGI, TensorRT-LLM, or tuned llama.cpp start to earn their complexity.
ca. $1,999 MSRP
Unlocks 11 additional models that do not fit on the current setup.
ca. $2,499 MSRP
RTX PRO 5000 Blackwell 48GBNVIDIA-UpgradeUnlocks 13 additional models that do not fit on the current setup.
ca. $4,999 MSRP
Unlocks 26 additional models that do not fit on the current setup.
ca. $2,499 MSRP
Unlocks 39 additional models that do not fit on the current setup.
Lifts average decode speed across fitting models by about 60%.
ca. $8,000 MSRP
RTX 5090 32GB (32 GB VRAM) can run these top models: Qwen3-Coder 30B A3B Instruct (score: 100/100), Qwen3-VL 30B A3B Instruct (score: 99/100), Qwen 3.5 27B (score: 98/100). See the full compatibility list above.
RTX 5090 32GB has 32 GB of VRAM available for AI model inference. This determines which models and quantization levels you can run locally.
Yes, RTX 5090 32GB is excellent for running LLMs locally with top compatibility scores above 80/100.
For coding on RTX 5090 32GB, we recommend Qwen 3.6 27B. It achieves 49.1 tokens per second with 187K context window. Qwen 3.6 27B is a specialized fit for Coding. It is a recent-generation family, which helps on current local SOTA workloads. It fits natively with comfortable headroom. Context coverage stays within the requested workload envelope. Known distribution channels: huggingface, lm-studio.
There are 5 upgrade path(s) from RTX 5090 32GB: RTX 5090 32GB, MacBook Pro M1 Max 64GB. Upgrading would unlock larger models and faster inference speeds.
Yes, RTX 5090 32GB with 32 GB of usable memory can run Flux.1 Dev at FP16 natively. Flux is a 12B parameter diffusion transformer that produces high-quality images. You can also run the Schnell variant for faster generation.
RTX 5090 32GB (32 GB VRAM) can handle various AI generation tasks beyond LLMs. For image generation, SDXL and Stable Diffusion 3.5 run well. Flux.1 Dev also runs natively for state-of-the-art image quality. For video, LTX Video 2.3 can generate short clips. Check the AI Capability Matrix above for detailed compatibility.
RTX 5090 32GB is excellent for AI image generation. With 32 GB of usable memory, it runs all major diffusion models including Flux.1, SDXL, and Stable Diffusion 3.5 at full precision. You can generate high-resolution images quickly and even handle video generation models.
Yes, RTX 5090 32GB with 32 GB of usable memory can run Qwen 3.5 27B at Q4_K_M (~16.5 GB) with ~7 GB headroom for context and runtime. Quality at Q4 is very close to full precision for most tasks. Run it with: ollama run qwen3.5:27b
With 32 GB on RTX 5090 32GB, Q4_K_M is the sweet spot for 27B-35B models, Q6_K for 14B models, and Q8_0 for 8B-9B models. The general rule: use the highest quantization that fits with at least 2-3 GB headroom for KV cache and runtime.
RTX 5090 32GB already has strong memory bandwidth, so the next limit is often memory capacity and context headroom rather than raw decode speed. For local LLMs, fit first and bandwidth second is the right mental model.
RTX 5090 32GB supports up to 2× GPU scaling via PCIe. With 2× GPUs, you get 64 GB effective memory with a 0.72× scaling factor per GPU. This enables running models like Qwen 3.5 397B A17B and Devstral 2 123B Instruct that don't fit on a single card.
RTX 5090 32GB uses PCIe for multi-GPU communication, which has approximately 28% scaling overhead. For best multi-GPU performance, consider NVLink-equipped variants.
Usually yes. If you want to run 2-4× RTX 5090 32GB for local AI, the bottleneck often becomes the platform, not the card. Workstation and server boards give you more CPU PCIe lanes, better x16 slot wiring, more spacing between cards, stronger power delivery, and usually more RAM capacity. Consumer x8/x8 layouts can work, but they are a common weak point in multi-GPU builds.
Compare with similar
Related guides
