What AI models can the RTX 5090 run?
The GeForce RTX 5090 has 32 GB of GDDR7 at 1,792 GB/s. With an 8K-token context it fits 19 of 21 models we track at Q4, 15 at Q8 and 11 at FP16; the largest text model it runs at Q4 is Mixtral 8x7B Instruct (MoE). The cheapest price we found today is $6,247 (used, GPU Sold, 2026-11-10), or you can rent one from $0.50/hr on Vast.ai.
Prices updated Nov 12, 2026 · CUDA - best support in llama.cpp, vLLM, ComfyUIBuy vs rent: at $6,247 to buy and $0.50/hr to rent, buying only pays off after about 12,569 GPU-hours (ignoring electricity and resale value). Run your own numbers →
Every model we track on the RTX 5090
| Model | Q4 | Q8 | FP16 |
|---|---|---|---|
| Mistral 7B Instruct v0.37.2B | ✓ ~196 tok/s6.1 GB | ✓ ~122 tok/s9.2 GB | ✓ ~69 tok/s15.5 GB |
| Llama 3.1 8B Instruct8.0B | ✓ ~180 tok/s6.5 GB | ✓ ~111 tok/s9.9 GB | ✓ ~62 tok/s17.0 GB |
| DeepSeek R1 Distill Llama 8B8.0B | ✓ ~180 tok/s6.5 GB | ✓ ~111 tok/s9.9 GB | ✓ ~62 tok/s17.0 GB |
| Qwen3 8B8.2B | ✓ ~174 tok/s6.7 GB | ✓ ~108 tok/s10.2 GB | ✓ ~61 tok/s17.4 GB |
| Gemma 3 12B IT12.2B | ✓ ~130 tok/s8.7 GB | ✓ ~77 tok/s13.9 GB | ✓ ~42 tok/s24.5 GB |
| Qwen3 14B14.8B | ✓ ~104 tok/s10.6 GB | ✓ ~63 tok/s16.9 GB | ✓ ~34 tok/s29.8 GB |
| DeepSeek R1 Distill Qwen 14B14.8B | ✓ ~101 tok/s10.8 GB | ✓ ~62 tok/s17.1 GB | ✓ ~34 tok/s30.0 GB |
| Mistral Small 24B Instruct (2501)23.6B | ✓ ~68 tok/s15.6 GB | ✓ ~40 tok/s25.6 GB | ✗needs 46 GB |
| Gemma 3 27B IT27.4B | ✓ ~60 tok/s17.5 GB | ✓ ~35 tok/s29.2 GB | ✗needs 53 GB |
| Qwen3 30B-A3B (MoE)30.5B | ✓ ~378 tok/s19.0 GB | ✓ ~246 tok/s32.0 GB | ✗needs 59 GB |
| Qwen3 32B32.8B | ✓ ~48 tok/s21.5 GB | ✗needs 35 GB | ✗needs 64 GB |
| Qwen2.5 Coder 32B Instruct32.8B | ✓ ~48 tok/s21.5 GB | ✗needs 35 GB | ✗needs 64 GB |
| DeepSeek R1 Distill Qwen 32B32.8B | ✓ ~48 tok/s21.5 GB | ✗needs 35 GB | ✗needs 64 GB |
| Mixtral 8x7B Instruct (MoE)46.7B | ✓ ~121 tok/s28.4 GB | ✗needs 48 GB | ✗needs 89 GB |
| Llama 3.3 70B Instruct70.6B | ✗needs 43 GB | ✗needs 73 GB | ✗needs 135 GB |
| Qwen3 235B-A22B (MoE)235.1B | ✗needs 135 GB | ✗needs 235 GB | ✗needs 440 GB |
| FLUX.1 [dev]11.9B · image | ✓ fits9.9 GB | ✓ fits15.0 GB | ✓ fits25.4 GB |
| FLUX.1 [schnell]11.9B · image | ✓ fits9.9 GB | ✓ fits15.0 GB | ✓ fits25.3 GB |
| Stable Diffusion XL 1.02.6B · image | ✓ fits5.2 GB | ✓ fits6.3 GB | ✓ fits8.6 GB |
| Wan 2.1 T2V 14B (video)14.3B · video | ✓ fits14.9 GB | ✓ fits20.9 GB | ✗needs 33 GB |
| Wan 2.1 T2V 1.3B (video)1.4B · video | ✓ fits4.6 GB | ✓ fits5.2 GB | ✓ fits6.4 GB |
✓ = fits in 32 GB with an 8K-token context (text models) and ~1 GB runtime overhead; tok/s is a rough single-user estimate (0.6 × bandwidth ÷ bytes per token), not a benchmark. Leave ~0.5–1 GB free if this card also drives your display. Method.
RTX 5090 head-to-head
Similar-VRAM cards, cheapest first
FAQ
How much VRAM does the RTX 5090 have?
32 GB of GDDR7 with 1,792 GB/s of memory bandwidth on a 512-bit bus.
What is the biggest AI model the RTX 5090 can run?
Of the models we track, Mixtral 8x7B Instruct (MoE) (47B) is the largest text model that fits at 4-bit (Q4) with an 8K context. Bigger models need a second GPU or CPU offload.
Can the RTX 5090 run Llama 3.3 70B Instruct?
Not on a single card: it needs about 43 GB at Q4 and the RTX 5090 has 32 GB.
Can the RTX 5090 run Qwen3 32B?
Yes, at Q4 (about 21 GB needed at Q4). Estimated speed at Q4: about 49 tokens/s.
Can the RTX 5090 run FLUX.1 [dev]?
Yes, at Q4, Q8, FP16 (about 10 GB needed at Q4).
How much does the RTX 5090 cost right now?
About $6,247 (used, GPU Sold, observed 2026-11-10). Prices change daily; check the retailer.
Is it cheaper to rent the RTX 5090?
For occasional use, yes. The cheapest on-demand rental we found was $0.50/hour (Vast.ai), so $6,247 buys about 12,569 hours of rental before electricity and resale value are considered.
Best GPU for…
Get the weekly GPU price drops
Cheapest card per model, new and used. Free, 1 email a week.
Subscribe on beehiiv →