What GPU do I need to run it?

How the VRAM and cost numbers are calculated

• Weight bytes = parameter count × bytes-per-parameter for the shown precision (FP32=4, BF16/FP16=2, FP8=1, INT4=0.5 bytes).
• Required VRAM adds a flat 1.2× overhead on top of weight size for activations and allocator fragmentation, plus KV-cache for text and text+image (VLM) models, calibrated for an assumed 8,192-token context. It's a rule-of-thumb, not a per-request simulation: long contexts or large batch sizes need materially more headroom for language models, diffusion/video models scale mainly with output resolution and frame count instead of context length, and speech models scale mainly with input audio length.
• “Cheapest live fit” only matches a GPU to a row if its hardware actually supports that row's compute format, and caps a fit at 8 GPUs of the same model. A pre-Ampere card (Pascal/Volta/Turing) never qualifies for a BF16 or FP8 row, regardless of VRAM or price. See the GPU Index for per-model specs.
• The INT4 row additionally requires a quantized checkpoint that's actually been published for the model in question. For language models this also means a GPU with AWQ/GPTQ/Marlin-class quantized-serving kernel support. Neither is guaranteed, so check the model's Hugging Face page for a published quantized variant before relying on that row.
• Prices are live per-GPU $/hr rates from Aquanode's marketplace feed, normalized the same way as the GPU Index, with outlier offers excluded.
• Multi-GPU fits assume the model shards cleanly across cards (e.g. tensor parallelism). Real throughput and interconnect overhead aren't modeled.