Best GPU for Local LLMs: Buyer's Guide by VRAM Tier

Short answer: buy VRAM first. A 16 GB card runs 14B-class models comfortably, 24 GB opens the 27B–35B class, and 32 GB (RTX 5090) is the ceiling for a single consumer card. Speed then follows memory bandwidth. Prices are unusually volatile in 2026, so this guide ranks cards by what they can run and shows launch MSRP and dated street price side by side.

Best GPUs by VRAM tier

Each tier lists the cheapest cards still sold new (by launch MSRP), the fastest one by memory bandwidth, and one used or end-of-life option where it matters. "Largest model that fits" is computed by the same engine as our GPU compatibility checker, so it matches what you will see there.

8–11 GB

Largest model that fits: GLM-4 9B (7.4 GB, Q4_K_M)

12–15 GB

Largest model that fits: Cosmos 3 Nano (11.9 GB, Q4_K_M)

16–23 GB

Largest model that fits: OLMo 2 13B Instruct (15.8 GB, Q4_K_M)

24–31 GB

Largest model that fits: Qwen 3.5 35B-A3B (23.8 GB, Q4_K_M)

32 GB+

Largest model that fits: Gemma 3 27B Instruct (25.4 GB, Q4_K_M)

Launch MSRP is the price at launch and never changes. Street price is what the card typically sells for, dated when it was checked; the 2026 memory shortage has pushed it far above MSRP on some cards.

How to choose

• VRAM decides what runs at all. A model that does not fit spills into system RAM and slows to a fraction of GPU speed. Sizes for every model and quantisation are in the VRAM requirements guide.
• Mixture-of-Experts models are fast for their size, not small. Every expert must be resident in memory, so a model such as Llama 4 Scout does not fit any single consumer card even though only 17B parameters are active per token.
• Memory bandwidth decides how fast it runs. Decode speed is set by how quickly the card can stream the active weights, which is why the bandwidth column matters as much as the VRAM one. Our methodology page publishes the formula.
• Power and size matter. The Power column is board power; check your power supply and case before ordering.

Prices in 2026: launch MSRP versus street price

Launch MSRP is what the manufacturer asked at launch and never changes. Street price is what a card actually costs now, and in 2026 the two have parted company because of the memory shortage. Figures below were checked on September 29, 2026 and will move:

• RTX 5060 Ti 16GB: about $805 new (median), roughly 88% above its $429 launch price (WCCFTech, TweakTown).
• RTX 5090: roughly $4,300 to $5,000 new against a $1,999 launch price, with some listings far higher (TechPowerUp, Tech Insider).
• RTX 50 Super: not on sale. NVIDIA has reportedly delayed the refresh indefinitely because 3 GB GDDR7 modules are scarce and costly, and reports on timing conflict, so do not plan a purchase around it (TweakTown, KitGuru).
• AMD Radeon RX 9000: AMD raised official prices during 2026 (WCCFTech, TweakTown).
• Used RTX 3090: price trackers disagree by about a third, so we do not publish a figure. Compare recent sold listings before you buy.

Rule of thumb: a card priced far above its launch MSRP is rarely the best value per gigabyte of VRAM. Compare the 16 GB and 24 GB tiers on price per GB before choosing.

Where to buy

These are the cards the tiers above point to. The links go to Amazon search results, so you see live prices and stock rather than a number we typed. We may earn a commission at no cost to you.

8 GB tier

16 GB tier

24 GB tier

32 GB tier

Unified memory and mini-PCs

Beyond a single GPU: unified memory and mini-PCs

If you want to run models bigger than 32 GB, a machine with a large unified memory pool can be the better route than stacking cards, at the cost of lower memory bandwidth per dollar than a fast discrete GPU.

• Apple Silicon: Apple announced Mac Studio with M5 Max and M5 Ultra in August 2026, with up to 512 GB of unified memory (9to5Mac, Apple). macOS caps how much of that memory the GPU may use by default, so check the usable share before you size a model.
• 128 GB mini-PCs: NVIDIA raised the DGX Spark Founders Edition to $4,699 in February 2026, while Strix Halo machines sit around $3,500 (OpenClaw DC, YUV.ai). See the AI stations and laptops listings for specs.

Frequently Asked Questions

How much VRAM do I need for local LLMs?

16 GB is the practical starting point: it runs 14B-class models at Q4 comfortably. 24 GB opens 27B to 35B models, and 32 GB adds headroom for longer context. Use the VRAM requirements guide to size the exact model you want.

Is a used RTX 3090 still worth it?

For 24 GB of VRAM at the lowest cost, usually yes, because 24 GB matters more than card age for local LLMs. Used prices are volatile right now, so compare recent sold listings and check the seller's return terms before you buy.

Should I wait for the RTX 50 Super?

No. It has been reported as delayed indefinitely, with conflicting timing, so it is not something to plan a purchase around. Buy for the VRAM tier you need today.

Is AMD or NVIDIA better for local LLMs?

Both work. NVIDIA has the broadest software support, so it is the lower-risk choice. AMD is worth comparing on VRAM per dollar. Check that your runtime supports the card before buying, and compare on VRAM and memory bandwidth first.

→ Check if your current GPU is enough | → VRAM Requirements Guide