Disclaimer

Read this before spending money based on a number from this site.

Estimates, not guarantees

Every figure produced by a Bitpute calculator is an estimate. The weights and KV cache portions are exact arithmetic from published data, but real-world memory use also includes framework buffers, CUDA context and fragmentation, which differ between driver versions, libraries and even individual runs. Our overhead term is an engineering estimate, informed by published memory-footprint numbers from llama.cpp and vLLM deployments and clearly labeled as such, yet your machine can and sometimes will differ. A result that says a model "fits" with a slim margin should be treated as a maybe, not a promise.

Prices and speeds change

Cloud rental prices, electricity rates and throughput figures shown in the workspace reflect the sources and dates listed on the LearnEngineering Center page. Providers change prices without notice. Always confirm current pricing with the provider before committing.

No affiliation

Bitpute is independent. We are not affiliated with, sponsored by, or endorsed by NVIDIA, AMD, Meta, Google, Alibaba, Mistral AI, Microsoft, DeepSeek, or any other hardware maker, model vendor or cloud provider named on this site. Their names appear purely to identify the products being calculated.

Not professional advice

Nothing on this site is financial, procurement or legal advice. For a purchase where being wrong is expensive, test on rented hardware first, the workspace can help you price exactly that.

Errors

If you find a figure that's wrong rather than merely different from your setup, please report it. Corrections are published in the release notes and the page's review date is updated.