For local AI, the amount of video memory (VRAM) matters more than anything else. A model that fits entirely on your graphics card runs fast. One that spills over into system RAM slows to a crawl. Using 4-bit models, 8GB of VRAM handles small models of about 8 billion parameters, 16GB handles models up to about 20 billion, and 24GB handles models up to about 32 billion. Our picks: the RTX 5060 Ti 16GB if you want the easiest setup, the RX 9070 16GB if you want the most speed for the money and don't mind a little setup, and a used RTX 3090 if you want 24GB for the bigger models.
We earn a commission from qualifying purchases made through links on this page, at no extra cost to you.
Our picks at a glance
| Easiest 16GB cardNvidia GeForce RTX 5060 Ti (16GB)First-time local AI users who want everything to work on the first try | 8.7 | Check price on Amazon | |
| Most speed for the moneyAMD Radeon RX 9070 (16GB)People comfortable with a little setup who also want a strong 1440p gaming card | 8.6 | Check price on Amazon | |
| Best for 27B to 32B modelsNvidia GeForce RTX 3090 (24GB, used or renewed)People who want to run 27B to 32B models fully on the GPU | 8.4 | Check price on Amazon | |
| The card I useAMD Radeon RX 7800 XT (16GB)AMD buyers on Windows who want official Ollama support | 8.2 | Check price on Amazon |
Why VRAM is the number that matters
When you run a model in a tool like Ollama or LM Studio, the whole model has to be loaded into memory. If it fits in your graphics card's VRAM, the GPU does all the work and answers come back quickly. If it doesn't fit, the leftover part runs from system RAM on your CPU, and speed drops sharply. It often runs several times slower.
The model isn't the only thing that uses VRAM. The conversation itself (the context) uses it too, and it grows as the chat gets longer or as you paste in documents. A model that just fits when you start can spill over once the conversation gets long. Leave a few gigabytes of headroom.
What fits on 8GB, 16GB and 24GB
These sizes assume 4-bit quantized models (the Q4 versions most people download), which keep most of a model's quality at about a quarter of the full size.
8GB: small models of about 7 to 8 billion parameters, like Llama 3.1 8B. These are good for quick questions, summaries and simple coding help, with short context.
16GB: models up to about 20 billion parameters, plus 8B to 14B models with plenty of room for long conversations. This is the sweet spot for most people.
24GB: models in the 27B to 32B range. A 27B model at Q4 is about 17GB on its own, so it doesn't fit on a 16GB card, but it runs fully on 24GB with room for context.
Bigger models like Llama 3.3 70B need about 42GB at Q4, which means two 24GB cards or a workstation card. Most people are better served by a strong 14B to 32B model.
Nvidia or AMD?
Nvidia is the easy choice. Nearly every local AI tool supports Nvidia's CUDA first, so things tend to work right after you install them, and new tools and features usually arrive for Nvidia before anyone else.
AMD works well in the popular tools. Ollama officially supports Radeon cards including the RX 7800 XT and RX 9070 series through AMD's ROCm drivers on Linux, and LM Studio runs on Radeon cards too. On Windows, Ollama's ROCm support list covers the RX 7000 series. The newer RX 9070 cards run there through the Vulkan backend instead, which works but is usually a little slower. AMD's advantage is more VRAM for the money.
If you're new to this and want the fewest surprises, buy Nvidia. If you're comfortable updating drivers and reading a setup guide, AMD gets you more card for the price.
-
#1Easiest 16GB card
Nvidia GeForce RTX 5060 Ti (16GB)
8.7Best for: First-time local AI users who want everything to work on the first try
Check price on AmazonThe RTX 5060 Ti 16GB is the cheapest current Nvidia card with 16GB of memory, which makes it the simplest way into local AI. Ollama, LM Studio, and image tools like ComfyUI all treat Nvidia as the default, so setup is usually just installing the app.
Prices are the problem. It launched at $429, but the memory shortage has squeezed supply. InsiderLLM reported eBay prices around $500 in August, and OpenClawDC saw new cards listed from about $740 to $1,099 in September. Check the current price before you buy.
Watch the listing title. There's also an 8GB RTX 5060 Ti that looks almost identical, and 8GB limits you to small models.
VRAM 16GB GDDR7 Launch price $429 Fits (4-bit) Models up to about 20B Software CUDA (works with nearly every tool) Interface PCIe 5.0 x8 What we like
- 16GB of VRAM, enough for models up to about 20B
- Nvidia CUDA support, so nearly every AI tool works without extra setup
- Low power draw for a 16GB card
Watch out for
- Stock is tight and street prices are well over the $429 launch price
- Make sure it's the 16GB model; the 8GB version is far less useful for AI
Verdict: The safest 16GB pick for local AI. Buy the 16GB model only, and compare its price with the RX 9070 first.
-
#2Most speed for the money
AMD Radeon RX 9070 (16GB)
8.6Best for: People comfortable with a little setup who also want a strong 1440p gaming card
Check price on AmazonThe RX 9070 has the same 16GB of memory as the RTX 5060 Ti and is a much faster card overall, so it's the better buy if you also game. It's our value pick in our GPU buy-now-or-wait guide for the same reason.
For AI, it runs in Ollama and LM Studio. On Linux, Ollama supports it directly through AMD's ROCm drivers. On Windows, it currently runs through the Vulkan backend, which works but is usually a bit slower than ROCm or CUDA.
The ASRock Challenger model we link was $749.99 on Amazon on October 31. If that's close to what an RTX 5060 Ti 16GB costs you, the RX 9070 gives you more card.
VRAM 16GB GDDR6 Launch price $549 Fits (4-bit) Models up to about 20B Software ROCm on Linux, Vulkan on Windows Amazon price (Oct 31) $749.99 (ASRock Challenger) What we like
- 16GB of VRAM
- Much faster than the RTX 5060 Ti in games
- Officially supported by Ollama on Linux
Watch out for
- Not on Ollama's Windows ROCm list yet, so it runs there through the slower Vulkan backend
- New AI tools usually support Nvidia first
Verdict: More speed than the 5060 Ti for gaming and AI, if you're fine with a little extra setup.
-
#3Best for 27B to 32B models
Nvidia GeForce RTX 3090 (24GB, used or renewed)
8.4Best for: People who want to run 27B to 32B models fully on the GPU
Check price on AmazonThe RTX 3090 is a 2020 card, but its 24GB of memory keeps it popular for local AI. That extra 8GB over a 16GB card is what lets 27B to 32B models run entirely on the GPU, which is a real step up in answer quality.
It isn't cheap anymore. OpenClawDC tracked an average of about $1,463 on eBay in September 2026, close to its original $1,499 price. At that price, it makes sense only if you'll actually run the bigger models.
Buy from a seller with a return window, and check that your power supply can handle it. Plan on 750 watts or more.
VRAM 24GB GDDR6X Original price $1,499 (2020) Used price seen (Sept 2026) About $1,450 (eBay average) Fits (4-bit) Models up to about 32B Power About 350W What we like
- 24GB of VRAM, enough for 27B to 32B models at 4-bit
- Fast memory, which matters a lot for how quickly answers come back
- CUDA support
Watch out for
- Only available used or renewed, so warranty coverage varies
- Used prices have climbed to about $1,450
- Uses about 350 watts and needs a strong power supply
Verdict: The most affordable way to get 24GB with CUDA. Buy it only if 16GB is holding you back.
-
#4The card I use
AMD Radeon RX 7800 XT (16GB)
8.2Best for: AMD buyers on Windows who want official Ollama support
Check price on AmazonThis is the card I run local models on. It has 16GB of memory, so it handles the same models as the other 16GB picks here.
Its edge is support. The RX 7800 XT is on Ollama's official ROCm list for both Windows and Linux, which the newer RX 9070 isn't on Windows yet. That makes it the smoother AMD choice for Windows users.
It launched at $499 in 2023. It's worth it if you find it noticeably cheaper than an RX 9070; otherwise buy the newer card.
VRAM 16GB GDDR6 Launch price $499 (2023) Fits (4-bit) Models up to about 20B Software ROCm on Windows and Linux What we like
- 16GB of VRAM
- On Ollama's official ROCm support list for both Windows and Linux
- Strong 1440p gaming card too
Watch out for
- Previous generation, so it can be harder to find new
- Slower than the RX 9070
Verdict: Proof that AMD works for local AI. A smart buy if it's priced below the RX 9070.
Frequently asked questions
How much VRAM do I need to run AI locally?
16GB for most people. Using 4-bit models, 8GB runs small models of about 8 billion parameters, 16GB runs models up to about 20 billion, and 24GB runs models up to about 32 billion. Leave a few gigabytes free for the conversation itself.
Can I run local AI on an AMD graphics card?
Yes. Ollama officially supports Radeon cards like the RX 7800 XT and RX 9070 through AMD's ROCm drivers on Linux, and LM Studio runs on Radeon cards too. Nvidia is still easier because most tools support it first.
What happens if a model is too big for my GPU?
The part that doesn't fit runs from system RAM on your CPU, and it gets much slower. Pick a smaller model or a smaller quantization so the whole thing fits on the card.
Is a used RTX 3090 worth it for local AI?
Only if you want to run 27B to 32B models. Its 24GB of memory fits them fully on the GPU, but used prices were about $1,450 in September 2026. If your models fit in 16GB, a newer 16GB card is cheaper and comes with a warranty.
How we picked
The RX 7800 XT is the card I run local models on, and the notes on what runs well come from my own use. The other picks come from research, not my own use. We ranked cards by VRAM first, since that decides which models run fully on the GPU, then by software support in Ollama and LM Studio, then by price. Model sizes are for 4-bit (Q4) versions. Prices change fast this year, so check the current listing before you buy.