GPU capacity and cost for open-weight LLMs
Use Calculate Compute to see what you're signing up for: how many GPUs, for how long, and at what cost. Every estimate shows its formula and its source.
Calculate Compute reproduces a validated spreadsheet model in your browser. Nothing you enter leaves this page.
Meta's Muse Glimmer-30B is loaded. Change any value, or paste another model's config.json and the fields fill themselves.
Training tokens and deadline, your fine-tuning dataset, or how many people you serve and how much context they use.
Summary gives the answer, the cheapest setup and the memory picture. AI Engineer explains every formula, what each symbol means, the numbers plugged in and the source.
Formulas follow Kundu et al. (2024) for training memory and serving, and Xia et al. (2024) for fine-tuning. Numbers in the AI Engineer view link to these sources.
Copy a plain-text summary of all three estimates, with the assumptions and prices used.