RTX 5090 Desktop (32 GB VRAM, 64 GB RAM) — Local LLM Performance

RTX 5090 Desktop (32 GB VRAM, 64 GB RAM) is a Desktop running Windows or Linux, with 32 GB VRAM + 64 GB system RAM.

Specifications

Memory for models	32 GB
Memory bandwidth	1,792 GB/s
Memory upgradeable	Yes
Chip	NVIDIA GeForce RTX 5090
Architecture	Blackwell GB202
Form factor	Desktop
Operating system	Windows or Linux
Released	2025
Llama 3.1 8B Q4_K_M	113–234 tok/s

Worth knowing before you buy

• 575 W board power — budget for a 1000 W+ PSU and the heat it puts into the room.
• Models larger than 32 GB must offload to system RAM, which costs roughly an order of magnitude in speed.

Models on the RTX 5090 desktop

• Yi 1.5 Family — Yes, but it is tight
• VibeThinker — Yes — comfortably
• Aya 3B (Tiny Aya) — Yes — comfortably
• StarCoder 2 — Yes — comfortably
• SmolLM3 — Yes — comfortably
• SmolLM2 — Yes — comfortably
• Qwen3-Coder — Yes — comfortably
• Qwen3.8 — Yes
• Qwen 3.7 — Yes, but it is tight
• Qwen 3.6 — Yes, but it is tight
• Qwen 3.5 — Yes, but it is tight
• Qwen 3 — Yes, but it is tight
• Qwen 2.5 VL — Yes — comfortably
• Qwen 2.5 Family — Yes, but it is tight
• Poolside Laguna XS 2.1 — Yes, but it is tight
• Phi-4 Mini — Yes — comfortably
• Phi-4 Family — Yes — comfortably
• Phi 3.5 Family — Yes — comfortably
• OLMo 2 — Yes — comfortably
• Nemotron Cascade 2 — Yes
• Nemotron 70B — Yes
• Nemotron 3 Nano Omni — Yes
• Mistral Small 3.1 — Yes
• Mistral Small 3.2 — Yes
• Mistral Family — Yes
• Ministral — Yes — comfortably
• Ministral 3 — Yes — comfortably
• Magistral Small — Yes
• Llama 3.3 — Yes
• Llama 3.2 Vision — Yes — comfortably
• Llama 3.2 Family — Yes — comfortably
• Llama 3.1 Family — Yes — comfortably
• InternLM 3 — Yes — comfortably
• IBM Granite 4.0 — Yes, but it is tight
• IBM Granite 4.1 — Yes
• Granite 3.0 — Yes — comfortably
• GPT-OSS — Yes — comfortably
• GLM-4.7 / GLM-Z1 — Yes, but it is tight
• GLM-6 — Yes — comfortably
• GLM-5 / GLM-5.1 — Yes, but it is tight
• Gemma 4 — Yes
• Gemma 3n — Yes — comfortably
• Gemma 3 — Yes, but it is tight
• Gemma 2 Family — Yes — comfortably
• Falcon 3 — Yes — comfortably
• EXAONE 3.5 — Yes, but it is tight
• Devstral — Yes
• DeepSeek R1 — Yes, but it is tight
• Cosmos 3 — Yes — comfortably
• Command R Family — Yes
• Cogito v1 — Yes, but it is tight
• Codestral — Yes
• Bonsai 27B — Yes, but it is tight
• BitNet b1.58 — Yes — comfortably
• Aya Expanse — Yes

← All hardware | AI stations and mini-PCs | Check your hardware