Best graphics cards for local AI under $500
Ranked by what each one runs fully in memory at 32k context, with prices in the US today.
In short
The top pick in the US right now is the NVIDIA Tesla P40, $405 at eBay: it runs Qwen 3.6 35B-A3B fully in memory at 32k context at about 50 tok/s (estimated). Ranked by the most capable open model each one runs fully in memory at 32k context and a usable speed, then by how many models it runs, then by speed on Qwen 3.8 27B, then by price.
Updated . Speeds are estimates; how they are worked out.
The list
Best first. Speeds are estimates est.
| # | Card | Memory | Price | Best model it runs at 32k | Models in memory | Qwen 3.8 27B | tok/s per $1,000 |
|---|---|---|---|---|---|---|---|
| 1 | NVIDIA Tesla P40 | 24 GB | $405eBay | Qwen 3.6 35B-A3BUD-Q4_K_M50 tok/s | 15 | 10 tok/s | 24 |
| 2 | Intel Arc A770 16GB | 16 GB | $400Amazon | gpt-oss 20BMXFP448 tok/s | 13 | offload | – |
| 3 | AMD Radeon RX 7600 XT | 16 GB | $480eBay | gpt-oss 20BMXFP429 tok/s | 4 | offload | – |
| 4 | Intel Arc B580 | 12 GB | $346eBay | Gemma 4 12BQ6_K21 tok/s | 10 | offload | – |
| 5 | NVIDIA GeForce RTX 3060 12GB | 12 GB | $360Amazon | Gemma 4 12BQ6_K21 tok/s | 10 | offload | – |
| 6 | NVIDIA GeForce RTX 3080 10GB | 10 GB | $469Amazon | Gemma 4 12BQ4_K_M59 tok/s | 5 | offload | – |
Questions
- What is the best graphics card for running AI models locally under $500?
- The top pick in the US right now is the NVIDIA Tesla P40, $405 at eBay: it runs Qwen 3.6 35B-A3B fully in memory at 32k context at about 50 tok/s (estimated).
- Which gives the most speed for the money?
- The NVIDIA Tesla P40: about 10 tok/s on Qwen 3.8 27B for $405, the most tokens per second per unit of price on this list (estimated).
- What is the cheapest option here that runs Qwen 3.8 27B fully in memory?
- The NVIDIA Tesla P40, $405 at eBay, at about 10 tok/s at 32k context (estimated).