Best used graphics cards for local AI

Ranked by what each one runs fully in memory at 32k context, with prices in Canada today.

In short

The top pick in Canada right now is the NVIDIA GeForce RTX 5090, $9,500 at eBay (used, asking price): it runs Qwen 3.8 27B fully in memory at 32k context at about 35 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. Used prices are eBay asking prices. Speeds are estimates est.

#CardMemoryPriceBest model it runs at 32kModels in memoryQwen 3.8 27Btok/s per $1,000
1NVIDIA GeForce RTX 509032 GB$9,500eBay, askingQwen 3.8 27BQ8_035 tok/s5235 tok/s3.7
2NVIDIA GeForce RTX 3090 Ti24 GB$3,000eBay, askingQwen 3.8 27BQ5_K_M27 tok/s4827 tok/s9
3NVIDIA GeForce RTX 409024 GB$4,450eBay, askingQwen 3.8 27BQ5_K_M27 tok/s4827 tok/s6
4NVIDIA GeForce RTX 309024 GB$2,149eBay, askingQwen 3.8 27BQ5_K_M25 tok/s4825 tok/s11.6
5AMD Radeon RX 7900 XTX24 GB$2,500eBay, askingQwen 3.8 27BQ5_K_M23 tok/s4823 tok/s9.1
6AMD Radeon AI PRO R970032 GB$2,800eBay, askingQwen 3.6 35B-A3BQ6_K64 tok/s3811 tok/s4
7NVIDIA GeForce RTX 4070 Ti Super16 GB$1,499eBay, askinggpt-oss 20BMXFP477 tok/s30offload–
8NVIDIA GeForce RTX 5070 Ti16 GB$1,600eBay, askinggpt-oss 20BMXFP4103 tok/s30offload–
9NVIDIA GeForce RTX 4080 Super16 GB$1,800eBay, askinggpt-oss 20BMXFP485 tok/s30offload–
10NVIDIA GeForce RTX 508016 GB$2,850eBay, askinggpt-oss 20BMXFP4110 tok/s30offload–
11AMD Radeon RX 7800 XT16 GB$950eBay, askinggpt-oss 20BMXFP464 tok/s29offload–
12AMD Radeon RX 9070 XT16 GB$1,000eBay, askinggpt-oss 20BMXFP465 tok/s29offload–

Questions

What is the best used graphics card for local AI in Canada?
The top pick in Canada right now is the NVIDIA GeForce RTX 5090, $9,500 at eBay (used, asking price): it runs Qwen 3.8 27B fully in memory at 32k context at about 35 tok/s (estimated).
Which gives the most speed for the money?
The NVIDIA GeForce RTX 3090: about 25 tok/s on Qwen 3.8 27B for $2,149 (used, asking price), 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 GeForce RTX 3090, $2,149 at eBay (used, asking price), at about 25 tok/s at 32k context (estimated).

Other lists