RTX 5090In stock
NVIDIA GeForce RTX 5090 Founders Edition
32GB card
$8,500at Newegg
1,792 GB/s (card only, needs a host PC)
Check priceMemory guide · 32GB VRAM
What can you run on 32GB VRAM?
8 of the 16 open models we track fit in 32GB VRAM without spilling into system RAM, the biggest being Qwen3.6 35B-A3B. Speeds range from about 73 to 269 tokens/sec depending on the model and machine. Prices & stock checked 2026-10-23. Prices change daily and some retailers block our checker, so always confirm on the store page.
Models that fit in 32GB VRAM
| Model | Needs at 4-bit | NVIDIA GeForce RTX 5090 Founders Edition1,792 GB/s |
|---|---|---|
| Qwen3.5 9Bdense · 9B active · chat, small agents | ~7 GB | Fits~135 tok/sQ8_0 |
| Gemma 4 12Bdense · 12B active · chat, vision | ~9 GB | Fits~104 tok/sQ8_0 |
| GPT-OSS 20BMoE · 3.6B active · reasoning, tools | ~13 GB | Fits~197 tok/sQ8_0 |
| Gemma 4 26B-A4BMoE · 3.8B active · fast general chat | ~18 GB | Fits~247 tok/sQ4_K_M |
| Qwen3.8 27Bdense · 27B active · coding, vision, general | ~19 GB | Fits~83 tok/sQ4_K_M |
| GLM-4.7-Flash 30B-A3BMoE · 3B active · coding agents | ~21 GB | Fits~269 tok/sQ4_K_M |
| Gemma 4 31Bdense · 31B active · general, vision | ~21 GB | Fits~73 tok/sQ4_K_M |
| Qwen3.6 35B-A3BMoE · 3B active · fast coding & chat | ~24 GB | Fits~269 tok/sQ4_K_M |
Each column shows the quant we'd use (8-bit when it fits comfortably and still runs 10+ tok/s, otherwise 4-bit) and the estimated decode speed for one user at ~8K context. How we estimate
Just out of reach: Llama 3.3 70B (~46 GB at 4-bit), Qwen3-Next 80B-A3B (~52 GB at 4-bit), Llama 4 Scout 109B-A17B (~71 GB at 4-bit).
1 tracked rig with 32GB VRAM
FAQ
What's the biggest model I can run on 32GB VRAM?
Of the models we track, Qwen3.6 35B-A3B is the largest that fits fully in 32GB VRAM (Q4_K_M, about 24 GB including runtime and ~8K context).
What's the cheapest rig with 32GB VRAM?
Of the machines we track, the NVIDIA GeForce RTX 5090 Founders Edition at $8,500 (Newegg, checked 2026-10-23) (card only, needs a host PC).
Can I run Llama 3.3 70B on 32GB VRAM?
Not fully in memory. At 4-bit it needs about 46 GB including runtime and ~8K context.
How are fit and speed estimated?
Memory needed = weights + 1.5 GB + 5% of weights. Decode speed ≈ 1 ÷ (active weights in GB ÷ (efficiency × memory bandwidth) + a small fixed per-token overhead), with efficiency and overhead fitted per platform to 43 public llama.cpp/MLX benchmarks (median error about 5%). They're estimates, not benchmarks.