Which Consumer GPU Should You Buy for Local AI in 2026?

Local text, image, and video models care about two things a gaming chart does not: how much VRAM (GPU memory) you have, and which number formats the chip can run in hardware. This guide is for a normal desktop with a discrete graphics card. It is aimed at people running lighter local models, not at a pure gaming build.

Unified-memory machines such as a Mac or an NVIDIA DGX Spark sit outside this guide. They are not discrete GPUs. They can split memory flexibly and hold a larger model than a small graphics card, and they are a reasonable way to run text. For image and video generation, a discrete NVIDIA GPU is still the more predictable choice.

Start with NVIDIA

On Windows and Linux, CUDA is still the path with the fewest dead ends in PyTorch, vLLM, ComfyUI, and Ollama. For that software stack:

1. Prefer a GeForce RTX 30-series card or newer. Newer architecture and more VRAM both help.
2. Skip old high-memory Tesla compute cards — the K, M, and P generations, such as the K80, M40, M60, P40, and P100. A large memory figure on those boards does not make up for the architecture.
3. Treat a Tesla V100 the same way for a new purchase. It is compute capability 7.0 (Volta). CUDA Toolkit 13.0 removed offline compilation and library support for Maxwell, Pascal, and Volta. Programs already built with CUDA 12 still run, and NVIDIA’s driver branch 580 is the last one for these GPUs, with support into 2028. Frameworks that move to CUDA 13 will not gain new builds for a V100. Our later tests show where a V100 we already had still wins, and where it loses.

Why newer usually beats “more memory”

Quantized models are stored in low-precision formats: FP16, BF16, FP8, FP4. When the GPU implements a format in hardware, it uses less memory and less time. When it does not, software converts the math to a format the chip does understand. The model can still run. It just spends extra memory and extra time to do it.

Native low-precision support on consumer GeForce cards, in practical terms:

Turing meets the compute-capability 7.5 line that many current tools expect, so a 20-series card is not in the “will not launch” group. It also has no native BF16, FP8, or FP4. That is why this guide starts the recommendation at the 30-series.

The video test later in this series is the concrete version of the table. The same LTX-2.3 workflow uses an FP8 checkpoint and an FP4 text encoder. A 16 GB RTX 5060 Ti held it in about 15 GB and finished faster than a 32 GB V100, which used about twice the memory. See RTX 5060 Ti vs Tesla V100 for LTX-2.3.

AMD on Linux, then everything else

AMD on Linux is the next tier. ROCm is a real stack there, and a Radeon card can run local models. At a similar price it is usually slower than NVIDIA, some attention kernels are less optimized, and a few cards are awkward to install. Plan on Linux if you buy AMD for this work.

Intel Arc, and AMD on Windows, can run jobs and still cost more evenings than they save. Intel’s PyTorch builds exist for Windows and Linux; in practice we have hit compatibility and stability problems. AMD’s Windows story is DirectML or community paths such as ZLUDA once you leave the officially supported recent generations, and those paths are slower to set up than CUDA. Buy one of these when the card is what you already own, or when the price gap is large and you enjoy fixing runtimes.

What to check, in order

VRAM first. It decides which model sizes load. For the models this guide is about, treat 16 GB as the minimum and 24 GB as the comfortable point. A 4-bit quant of a 12B model fits in 16 GB and runs well. A 4-bit 27B model does not, and the card then shares the work with the CPU. We measured both cases on an RTX 5060 Ti 16GB: Gemma 4 12B and Qwen3.6-27B.

Architecture second. Compute capability 7.5 or newer (Turing and later) is the line for mainstream CUDA frameworks such as vLLM and ComfyUI. Among cards that clear it, 30-series is usable, 40-series is better, and 50-series is the one with native FP4.

The tables below mark whether a card clears the 16 GB line. They are not a full catalog. An 8 GB or 12 GB card missing from a list fails that line too.

GeForce RTX 50-series (Blackwell)

The RTX 5060 Ti 16GB is the card we actually tested. It is a PCIe x8, 128-bit part, which looks weak next to a 5090 and next to old compute cards with huge memory buses. For models that fit, the newer tensor cores mattered more than that bus. The write-up of the card itself is RTX 5060 Ti 16GB for local AI.

RTX PRO Blackwell workstation cards

Memory sizes below match NVIDIA’s workstation specifications. All of these clear 16 GB.

The 72 GB 5000, the 5500, and the Max-Q 6000 were not in our original June 2026 table. They are on NVIDIA’s page as of this update. Read that page before you order a workstation card; the lineup has been moving.

GeForce RTX 40-series (Ada Lovelace)

A stock RTX 4080 SUPER is a 16 GB card. Extra memory on one of these boards is an aftermarket modification, which is a different product from the card in this table.

RTX Ada workstation cards

GeForce RTX 30-series (Ampere)

A used RTX 3090 is still the 30-series card that makes sense for local models, because of the 24 GB. The 10 GB and 12 GB cards in this generation miss the line even when the GPU itself is fast.

RTX A-series workstation cards

What we would buy

For a desktop that runs local chat, images, and the occasional video model:

• Buy NVIDIA, 30-series or newer, with at least 16 GB. Prefer 24 GB when the price is close.
• Prefer Ada if you can, and Blackwell if the card that fits your budget has it. Native FP8 and FP4 are the reason, not the generation number on the box.
• On a tight budget, an RTX 5060 Ti 16GB is the Blackwell card that clears the memory line. The 8 GB 5060 Ti does not. We paid about ¥3,500 for a 16 GB card in June 2026 and would hesitate above ¥4,000. The tests are linked above.
• Do not buy a cheap old Tesla, including a V100, as if it were a 2026 gaming GPU with extra memory. Use one you already have while CUDA 12 builds still run. Do not make it the card you depend on for new software.