Two cards give you 32 GB of VRAM, the memory on the graphics card: the GeForce RTX 5090 and the RTX PRO 4500 Blackwell. Same capacity means the same list of models fits. This post is that list, with the speed each card is estimated to deliver.
We already explained how to work out the memory a model needs in our guide to GPU memory for local LLMs. Here we skip the formula and use the answer: the selector counts 31.4 GB as usable on a 32 GB card, because the driver keeps a little for itself.
The model list
All rows run fully in VRAM at Q4 (a 4-bit model file, about a third of the original size) with a 32,000-token context, which is the working memory for roughly 24,000 words. The one exception is gpt-oss-20b, shown at Q8 because it is small enough. Speed is tok/s, tokens per second, roughly words per second, for one user.
| Model | Good for | Memory | RTX 5090 | RTX PRO 4500 |
|---|---|---|---|---|
| Qwen3.8 27B | All-round assistant | 21.5 GB | about 63 tok/s | about 35 tok/s |
| Gemma 4 31B | Writing, analysis | 24.9 GB | about 55 tok/s | about 30 tok/s |
| Mistral Small 3.2 24B | Drafting, summaries | 21.3 GB | about 62 tok/s | about 35 tok/s |
| Gemma 4 26B-A4B | Fast assistant | 19.6 GB | about 188 tok/s | about 132 tok/s |
| Qwen3-Coder 30B-A3B | Vibe-coding | 23.5 GB | about 153 tok/s | about 100 tok/s |
| Nemotron SEA-LION 30B-A3B | Indonesian and regional languages | 27.6 GB | about 200 tok/s | about 145 tok/s |
| gpt-oss-20b (Q8) | Light jobs, many users | 14.5 GB | about 200 tok/s | about 145 tok/s |
Vibe-coding means describing what you want in plain words and letting the model write the code. A coder model such as Qwen3-Coder is built for that.
Notice the two groups. The dense models (Qwen3.8, Gemma 4 31B, Mistral Small) use every weight for every word, so they run at 30 to 63 tok/s. The MoE models (mixture-of-experts: many weights, but only a few billion active per word) run at 100 to 200 tok/s. Both groups are faster than most people read, which is about 5 to 10 tok/s.
Estimated speed on a 32 GB card, 4-bit model, 32K context
Estimate. Tokens per second for one user. Higher is faster.
Source: EMD Lab AI selector estimate, calculated from model files and card specifications. Not a measurement.
What does not fit
Models above about 30B in size break the 32 GB limit. The selector shows what happens when part of the model sits in system RAM (offload):
- Llama 3.3 70B at Q4 needs about 55.7 GB. With 57% of it offloaded, the estimate is about 1.5 tok/s on either card. That is too slow to use.
- gpt-oss-120b needs about 67 GB, but it is an MoE model, so offload hurts less: about 20 tok/s on either card. Usable for one person, not a team.
- Q8 versions of the 27B to 31B dense models need 33 to 38 GB. They run with a small offload, at an estimated 5 to 12 tok/s. Stay at Q4.
Past the point where the model no longer fits, the two cards look the same, because the slow part is system RAM, not the card. For 70B and larger, look at a card with 96 GB.
Users matter too
The 21.5 GB figure for Qwen3.8 27B is for one user. Each extra person adds their own context. At 32K the selector estimates room for about 5 conversations at once on a 32 GB card, and only about 2 for Gemma 4 31B. A 32 GB card is a one-person or small-team card. For a shared box, see the 96 GB cards.
What else a 32 GB card does
A chat model is not the only job. These estimates come from the selector's workload table, and they are rough:
- Speech to text: Whisper large-v3-turbo needs about 4 GB comfortably. It handles Indonesian.
- Embeddings for RAG search: bge-m3 needs about 4 GB. A larger embedding model, Qwen3-Embedding-8B, wants about 24 GB for large batches.
- Image generation: Qwen-Image fits at about 20 GB. FLUX.2 [dev] lists about 32 GB as comfortable, so it sits right at the limit.
- Video generation: Wan 2.2 and LTX-2.5 list about 32 GB as comfortable.
Check the licence of any image or video model before commercial use; some are free only for non-commercial work or below a company-size limit.
Same capacity, different card
| GeForce RTX 5090 | RTX PRO 4500 Blackwell | |
|---|---|---|
| Memory | 32 GB GDDR7 | 32 GB GDDR7 with ECC |
| Bandwidth | 1,792 GB/s | 896 GB/s |
| Power | 575 W | 200 W |
ECC means error-correcting memory, which catches and fixes rare bit errors. The 5090 reads memory twice as fast, so dense models run close to twice as fast on it, as the table above shows. The PRO 4500 draws a third of the power, which matters for a small office on a limited PLN supply, and it has ECC and workstation drivers.
Who picks which
RTX 5090: one or two heavy users, coding assistants, a dense 27B or 31B model where you wait for each answer, or image and video work where speed counts. It needs a large power supply and good airflow.
RTX PRO 4500: an office assistant, document questions, meeting minutes and drafting, where 30 to 35 tok/s is already faster than reading. It is the card in the standard Studio build, and it suits a machine that runs all day.
The short version
- 32 GB runs 27B to 31B dense models at Q4, and 30B MoE models, fully in VRAM.
- MoE models are the fast choice: 100 to 200 tok/s estimated. Dense models give 30 to 63.
- 70B does not fit (about 1.5 tok/s). gpt-oss-120b runs with offload at about 20 tok/s.
- The 5090 is about twice as fast on dense models; the PRO 4500 uses 200 W instead of 575 W and has ECC.
- For many users at once, or 70B and up, go to 96 GB.
All speeds and memory figures are estimates calculated by the selector. Try your own model in the model finder, or open the standard Studio build.