Best graphics cards for local AI
For running AI models at home, video memory beats everything. It decides which models fit on the card, and memory bandwidth decides how fast they talk back. These are the cards worth buying for Ollama, LM Studio and llama.cpp at today's UK prices, new and used.
Check what your PC can run → Pick your card and RAM, see which Qwen, DeepSeek and Llama models run and how fast.
-
Cheapest 24GB
1. NVIDIA RTX 3090 24GB (used)
Used on eBay · 24GB GDDR6X · 936GB/s
The cheapest way to 24GB: it holds 30B-class models like Qwen3 32B and Qwen3-Coder entirely on the card, with 936GB/s memory close to a 5080's. AI buyers have pushed used prices to around £1,000, about the same as a new 5070 Ti, so buy it for the extra 8GB and pick a seller who offers returns.
-
Best new card for local AI
2. ASUS Prime RTX 5070 Ti OC 16GB
RTX 5070 Ti · 16GB GDDR7
16GB and 896GB/s with CUDA. Runs every 14B model and gpt-oss 20B on the card, and Qwen3 30B-A3B quickly with RAM helping.
Used on eBay right now
-
Fastest 16GB
3. Gigabyte RTX 5080 Gaming OC 16GB
RTX 5080 · 16GB GDDR7
Same 16GB as the 5070 Ti with a little more bandwidth, so slightly quicker replies. Worth it if you also game at 4K. For AI alone, the 5070 Ti is better value.
Used on eBay right now
-
Cheapest new 16GB
4. PowerColor Reaper RX 9060 XT 16GB
RX 9060 XT · 16GB GDDR6
The lowest price for 16GB new. Its 320GB/s memory makes it about a third as fast per word as a 5080, and you'll use LM Studio or Ollama's AMD build rather than CUDA tools.
Used on eBay right now
-
Most VRAM you can buy new
5. ASUS ROG Astral RTX 5090 OC 32GB
RTX 5090 · 32GB GDDR7
32GB runs every 32B-class model on the card with room for long conversations. At this price it's for people billing for AI work.
-
Fastest 24GB
6. NVIDIA RTX 4090 24GB (used)
Used on eBay · 24GB GDDR6X · 1,008GB/s
Same 24GB as a 3090 with about 8% more bandwidth and far quicker image generation. Used ones start around £1,800, so for text models a 3090 is the better buy.
-
Fast AMD option
7. PowerColor Reaper RX 9070 XT 16GB
RX 9070 XT · 16GB GDDR6
640GB/s and 16GB for less than a 5070 Ti. Good with LM Studio and llama.cpp; check your tools support Radeon (ROCm or Vulkan) first.
-
Right card, wrong price
8. ASUS Prime RTX 5060 Ti OC 16GB
RTX 5060 Ti · 16GB GDDR7
16GB with CUDA is exactly what local AI wants, but at this price a used RTX 3090 is worth the extra £200 or so: 24GB and twice the bandwidth. Set a watch.
Used on eBay right now
-
Cheapest way in
9. Sparkle Intel Arc B580 Titan OC 12GB
Arc B580 · 12GB GDDR6
12GB for under £350 runs 8B and 14B models on the card. Support comes through llama.cpp's Vulkan backend and LM Studio, so expect some setup.
Used on eBay right now
-
12GB holds you back
10. MSI RTX 5070 Gaming Trio OC 12GB
RTX 5070 · 12GB GDDR7
Fast memory, but 12GB stops at 14B models on the card. For AI, spend the extra on a 16GB card.
Used on eBay right now
Nothing matches those filters. Try fewer filters, or search for a brand like "MSI" or "Crucial".
How we rank
Ranked on video memory per pound, memory bandwidth (which sets words per second), and how easy the software is: NVIDIA's CUDA just works, AMD and Intel need a little setup. Model sizes and speeds checked against what people are actually searching for: Qwen, DeepSeek, Gemma and gpt-oss. Prices checked daily.
Prices are Amazon UK and eBay UK at the time we last checked; they move daily, so confirm on the retailer's page. We earn a small commission if you buy through our links, at no cost to you. It never changes the order. More about how we work.