Best graphics cards for running AI models locally (2026)

A graphics card is the fastest way to run AI models on a desktop, as long as the model fits in its video memory. Our own 8 GB RTX 4060 writes Qwen3.5 9B at 45.5 tokens per second (test T0002), far above reading speed, but it cannot hold anything much bigger. So the question for a new card is simple: how much video memory can you get, and how fast is it?

2026 makes that harder. The memory these cards use is scarce, and in September 2026 most cards sold far above their launch prices, the RTX 5060 Ti for nearly double. We picked the cards that give the most local AI for the money at the prices we found at Newegg and Best Buy on September 26, 2026, from all three makers.

What each card can run

Estimated writing speed, tokens per second

Machine	Qwen3.5 9B5.68 GB file	gpt-oss 20B12.11 GB file	Qwen3.8 27B16.46 GB file	Qwen3.6 35B-A3B22.13 GB file
XFX Swift Radeon RX 9060 XT 16GB 16 GB, 320 GB/s	40-52	54-70	6-8*	29-39*
ASUS Prime GeForce RTX 5060 Ti 16GB 16 GB, 448 GB/s	61-79	111-144	8-11*	53-73*
ASRock Intel Arc Pro B60 Creator 24GB 24 GB, 456 GB/s	43-56	34-44	15-19	33-43
ASRock Radeon AI PRO R9700 Creator 32GB 32 GB, 640 GB/s	76-98	98-127	27-34	97-125

These estimates use the same method we validated on our own hardware and against published results, plus a correction for AMD and Intel cards: in published measurements their software gets less speed out of the same memory bandwidth than NVIDIA’s, and Intel’s falls furthest behind on mixture-of-experts models such as gpt-oss and Qwen3.6 35B-A3B.

The picks

Best for most people Researched, not tested

XFX Swift Radeon RX 9060 XT 16GB

16 GB GDDR6, dual fan (RX-96TSW16BQ). AMD Radeon RX 9060 XT (RDNA 4)

Video memory

16 GB

Bandwidth

320 GB/s

Price

$549.99

The least expensive 16 GB graphics card we found in September 2026. At 320 GB/s it runs models up to about 14B, plus gpt-oss 20B, entirely on the card at well above reading speed, though more slowly than an RTX 5060 Ti.

Sixteen gigabytes is the sweet spot for adding local AI to an existing PC: it holds models up to about 14B entirely on the card, plus gpt-oss 20B, and bigger mixture-of-experts models run with part of the model in system RAM. In September 2026 the RX 9060 XT was the cheapest way to get it, at $549.99, about $230 less than the cheapest RTX 5060 Ti we found. It is not the fastest 16 GB card, but an estimated 54 to 70 tokens per second on gpt-oss 20B is still several times reading speed. LM Studio runs it out of the box with its Vulkan engine; some other tools need AMD’s ROCm software and a little more setup.

What we like

• 16 GB for about $230 less than the cheapest RTX 5060 Ti we found
• Runs gpt-oss 20B at an estimated 54 to 70 tokens per second
• 160 W typical board power; AMD recommends a 450 W power supply

What we don't

• Slower than an RTX 5060 Ti, especially on mixture-of-experts models
• Some local AI tools need more setup on AMD than on NVIDIA

Where to buy

• Check price at Newegg 16 GB, Sold by XFX
• Check price on Amazon 16 GB

$549.99 at Newegg (sold by XFX) for 16 GB on Sep 26, 2026; list price $349 (AMD launch price for the 16 GB model, June 2025; checked Sep 26, 2026). Prices change often; the buttons show today's price at each store.

Easiest to use Researched, not tested

ASUS Prime GeForce RTX 5060 Ti 16GB

16 GB GDDR7. NVIDIA GeForce RTX 5060 Ti (Blackwell)

Video memory

16 GB

Bandwidth

448 GB/s

Price

$829.99

The least expensive way to get 16 GB of NVIDIA video memory, though in September 2026 it sold for nearly double its launch price. NVIDIA's CUDA software works with every local AI app, and at 448 GB/s it runs models up to about 14B parameters, plus gpt-oss 20B, entirely on the card.

If you want no software hassle, NVIDIA is still the safe choice: its CUDA software works with every local AI app, and by our estimates the RTX 5060 Ti is about twice as fast as the RX 9060 XT on mixture-of-experts models such as gpt-oss 20B. The catch is price. It launched at $429, and in September 2026 the 16 GB version cost $779.99 to $829.99 at Newegg. For about the same money, AMD’s RX 9070 XT (16 GB, 640 GB/s) is faster on dense models and slightly slower on mixture-of-experts models, by our estimates, with the same software caveats as the RX 9060 XT.

What we like

• CUDA works with every local AI app, no workarounds
• 16 GB fits models up to about 14B and gpt-oss 20B on the card
• About twice as fast as the RX 9060 XT on mixture-of-experts models, by our estimates

What we don't

• Nearly double its $429 launch price in September 2026
• 128-bit memory bus limits speed compared with bigger cards

Where to buy

• Check price at Newegg 16 GB, OC edition
• Check price on Amazon 16 GB

$829.99 at Newegg for 16 GB, OC edition on Sep 26, 2026; list price $429 (NVIDIA launch price for the 16 GB model, April 2025; checked Sep 26, 2026). Prices change often; the buttons show today's price at each store.

Most memory per dollar Researched, not tested

ASRock Intel Arc Pro B60 Creator 24GB

24 GB GDDR6, blower cooler (B60 CT 24G). Intel Arc Pro B60 (20 Xe-cores, Xe2)

Video memory

24 GB

Bandwidth

456 GB/s

Price

$649.99

The cheapest way to get 24 GB of video memory on a new card, for less than a 16 GB NVIDIA card in September 2026. Intel's software for local AI is still catching up, so it runs dense models well but mixture-of-experts models far slower than its hardware allows.

Twenty-four gigabytes for less than a 16 GB NVIDIA card is a real advantage: at $649.99 the B60 fits dense 27B models that won’t fit on 16 GB, at an estimated 15 to 19 tokens per second. The trade-off is software. Intel’s cards need its SYCL build of llama.cpp for their best results, and mixture-of-experts models currently run at a fraction of the speed the hardware should manage. Choose it if your models of choice are dense, and you don’t mind a little setup.

What we like

• 24 GB of video memory for less than a 16 GB NVIDIA card
• Fits dense 27B models that don't fit on 16 GB cards
• Blower cooler exhausts heat out of the case, good for multi-card builds

What we don't

• Local AI software support lags NVIDIA and AMD, especially for mixture-of-experts models
• Best results need Intel's SYCL build of llama.cpp rather than the default

Where to buy

• Check price at Newegg 24 GB, blower
• Check price on Amazon 24 GB, blower

$649.99 at Newegg for 24 GB on Sep 26, 2026. Prices change often; the buttons show today's price at each store.

Best 32 GB card Researched, not tested

ASRock Radeon AI PRO R9700 Creator 32GB

32 GB GDDR6, blower cooler. AMD Radeon AI PRO R9700 (64 compute units, RDNA 4)

Video memory

32 GB

Bandwidth

640 GB/s

Price

$1,699.99

A 32 GB workstation card built on AMD's RDNA 4 chip. It has as much video memory as an RTX 5090 for about a quarter of that card's September 2026 price, with 640 GB/s of bandwidth and good support in llama.cpp and LM Studio.

Thirty-two gigabytes is where a single card gets comfortable with the popular 27B to 35B models, including long conversations. The R9700 has the same amount of memory as an RTX 5090 with about a third of its bandwidth, for about a quarter of its September 2026 price: $1,699.99 at Newegg. Published results put it at about 30 tokens per second on a dense 27B model and over 120 on a 35B mixture-of-experts model. It reads long prompts slower than an NVIDIA card would.

What we like

• 32 GB of video memory, enough for 27B to 35B models on one card
• 640 GB/s bandwidth
• Blower cooler suits workstation and multi-card builds

What we don't

• About $400 above its $1,299 launch price in September 2026
• Reads long prompts slower than comparable NVIDIA cards
• Some tools need extra setup compared with CUDA

Where to buy

• Check price at Newegg 32 GB, blower
• Check price on Amazon 32 GB, blower

$1,699.99 at Newegg for 32 GB, blower on Sep 26, 2026; list price $1,299 (AMD retail launch price, October 2025; checked Sep 26, 2026). Prices change often; the buttons show today's price at each store.

What about Intel’s Arc Pro B70?

Intel’s 32 GB Arc Pro B70 (608 GB/s) launched in March 2026 at $949, well below the R9700. In September 2026 we found it for $1,894.99 to $2,120.25 from US sellers, more than the R9700, which by our estimates is faster on both dense and mixture-of-experts models, so we don’t recommend it at current prices. Intel’s Arc Pro B65 (also 32 GB and 608 GB/s) is meant to cost less than the B70, but we found it only on Amazon, where we can’t quote prices; it has the same software caveats as the B60.

What about the RTX 5090?

It is the fastest consumer card for local AI, with 32 GB and 1,792 GB/s. But the cheapest one we found at Newegg in September 2026 cost $6,885.90, more than three times its $1,999 list price, and for less money a 128 GB unified-memory mini PC or Mac holds four times the model size. Unless you need the fastest possible speed on models that fit in 32 GB, we would wait for prices to fall.

What about a used RTX 3090?

With 24 GB of video memory and 936 GB/s of bandwidth, the RTX 3090 from 2020 remains one of the best local AI cards per dollar on the used market, and it runs NVIDIA’s software. Buy from a seller with a return policy, check the card’s temperatures and fans on arrival, and make sure your power supply can handle 350 W.

A graphics card or a unified-memory computer?

A graphics card wins on speed for models that fit, and it upgrades a PC you already own. A unified-memory machine wins on size: a 128 GB Ryzen AI Max+ 395 mini PC or Mac holds 70B and 120B-class models that no single consumer card can. Our mini PC guide covers those, and our calculator shows what fits on either.

How we chose

We compared current cards with 16 GB or more of video memory on memory size, bandwidth and the prices we found at Newegg and Best Buy on September 26, 2026, and estimated their speed with the method in our bandwidth study, corrected with published measurements for AMD and Intel cards. We have not tested these cards ourselves yet; our own graphics card results are from an RTX 4060.