Choose what you want to do, compare a few cards, then check the exact listing. For desktop upgrades in the US. Prices cover the graphics card only.
Amazon US · USD · Recorded offers, not live checkout prices. Each quote has its own date. Fresh quotes are preferred; older quotes are clearly identified. New and Used prices can come from different dates. As an Amazon Associate I earn from qualifying purchases.
Try it first
Use your existing computer
You do not need to buy a graphics card just to install a local AI app. Try a small model first; your system still needs enough memory, and CPU generation can be slower.
Start with Ollama and a small Qwen3 4B model. On a compatible computer with enough system memory, CPU or supported integrated/unified-memory hardware may be enough to explore. You can decide whether to upgrade after trying it.
Everyday chat
Start by comparing 12–16GB cards
A starting point for small, compressed chat models, such as an 8B model at 4-bit precision. More memory leaves room for the conversation and the app.
Qwen3 8B Q4_K_M is a 5.2GB download. That is the model file, not its total working memory. Model package sizes.
These are planning ranges, not guaranteed fit or speed. Adjust the 8B memory example for your model and conversation length.
RTX 3060 · 12GB
A 12GB starting option. Confirm the 12GB version, rather than the 8GB version.
$399.99 · Used Last recorded quote
Quote dated 2026-10-05 08:57:27 UTC. Older quote; confirm today’s price and stock before deciding.
No recorded offers match these choices. Widen the budget or select both New and Used.
Coding & more room
Compare 16GB and 24GB cards
More room to experiment with a 14B compressed model or keep more conversation in memory. Coding is also possible with smaller models; long documents and coding agents can need much more memory.
Qwen3 14B Q4_K_M is a 9.3GB download, before conversation memory and runtime overhead. Model package sizes.
These are planning ranges, not guaranteed fit or speed. Adjust the 14B memory example for your model and conversation length.
RTX 5060 Ti · 16GB
Another 16GB option. Compare today’s price with the 4060 Ti; this guide does not rank their speed.
$788.50 · New Last recorded quote
Quote dated 2026-09-29 14:05:39 UTC. Older quote; confirm today’s price and stock before deciding.
No recorded offers match these choices. Widen the budget or select both New and Used.
Larger models
Compare 24GB and 32GB cards
A starting point for experimenting with compressed 30–32B models. A 24GB card has less room for conversation memory than a 32GB card; check the exact model and context before buying.
Qwen3 32B Q4_K_M is a 20GB download. Its 8-bit version is 35GB before overhead, so the two versions need different hardware. Model package sizes.
These are planning ranges, not guaranteed fit or speed. Adjust the 32B memory example for your model and conversation length.
RTX 3090 · 24GB
A 24GB option, including recorded Used offers. Check power, case space and return terms.
$2,195.00 · New Last recorded quote
Quote dated 2026-10-07 06:17:19 UTC. Older quote; confirm today’s price and stock before deciding.
No recorded offers match these choices. Widen the budget or select both New and Used.
The terms, in plain English
VRAM
The card’s memory. More room helps hold the model and conversation; it does not guarantee a faster or smarter model.
8B / 14B / 32B
The approximate model size in billions of parameters. Larger models usually need more memory, but size alone does not measure quality.
4-bit / Q4
A compressed version of model weights that uses less memory, with possible quality tradeoffs. Compare the exact model version.
Context
How much text the model can keep track of at once. Longer chats, documents and multiple users can need more memory.
Before you buy
Confirm the exact memory version. RTX 4060 Ti and RTX 5060 Ti come in both 8GB and 16GB versions.
Check your case space, power supply and power connectors against the exact card manufacturer’s specifications. This is a card upgrade, not a complete computer.
Check your chosen app’s GPU, operating-system and driver support. This shortlist uses NVIDIA’s documented CUDA path; AMD, Intel and Apple can also work with supported backends.
For Used cards, inspect the listing condition and return terms. We have not tested these individual cards.
Check Amazon’s current price and availability. A recorded quote is not a promise of today’s price or a claim of the best deal.
What about 70B models or two GPUs?
A dense 70B model at ideal 4-bit precision needs about 32.6GiB for weights alone, before compression metadata, conversation memory and app overhead. A 32GB card is not a safe all-on-GPU recommendation for that case. Multi-GPU setups and CPU offloading depend on the runtime and are beyond this beginner single-card shortlist. Use the VRAM calculator to plan allowances.
Why these cards and categories?
These are editorial starting choices, selected for documented NVIDIA runtime support and exact memory variants. They are not benchmark rankings, Amazon best-seller recommendations or claims to outperform every alternative. We have not measured tokens per second. Cards requiring external enclosures, special BTF motherboards or liquid-cooling setups are excluded from the shortlist.
For each exact card and condition, we show the lowest qualifying fresh quote when available. Otherwise, we show the lowest recorded quote from the most recent quote date in our 30-day sample. These are selected catalog offers, not market-wide lowest prices.
The memory bands are our planning interpretation of published model package sizes plus runtime/context overhead. The calculator’s 2GiB conversation-cache, 1GiB runtime and 20% headroom examples are adjustable illustrations, not measured requirements.