For local models, VRAM decides what runs and memory bandwidth decides how fast. These are the 13 consumer GPUs in our table with a US launch MSRP of $500 or less, scored on both with the same formulas as the rest of the site. Prices are launch MSRPs, not today's street prices, which can be higher or lower.
RX 7800 XTmost VRAM (16 GB, $499)
RX 7800 XTfastest (624 GB/s, $499)
167models at 4-bit+ on the Arc A770 16GB
Arc B580cheapest listed ($249)
The short answer
Run the biggest models: the RX 7800 XT has 16 GB, and runs 167 of 298 open models we track fully in VRAM at 4-bit or better (8K context). The largest popular one is gpt-oss 20B (21B).
Lowest launch price: the Arc B580 ($249, 12 GB) runs 150 models at 4-bit+.
Every card under $500
Sorted by VRAM, then bandwidth. Model columns show the best quant that fits fully in VRAM at 8K context and an estimated speed ceiling, or the llama.cpp offload plan when the model only runs with part of it in system RAM (speed range from DDR4-3200 to DDR5-6000). How we calculate · offload method.
Start from the model you want. A model needs roughly its 4-bit file size plus KV cache and about 1 GB of runtime overhead. Pick the cheapest card where it fits with room for your context length, then prefer higher bandwidth for faster replies. If a model is just out of reach, mixture-of-experts models such as gpt-oss and Qwen3 A3B can still run with most expert weights in system RAM at a usable speed; dense 70B models usually cannot.
Prices: NVIDIA, AMD and Intel US launch MSRPs (Founders/reference editions). Specs: vendor spec pages, linked on each GPU page. Speed is a bandwidth ceiling, not a benchmark; real speed is typically 50–80% of it. Data refreshes hourly. Report an error.