LOCAL MEMORY BUDGETMATH.JS 13.2.0 BIGNUMBER READY

Know what fits before the model loads.

Separate model weights from runtime and context memory. Every input is an assumption; every result is exact decimal arithmetic.

Host and model assumptions
FIT
Memory headroom29.232 GB
35 + 7 + 32.768 = 74.768 GB model load
Usable104 GB
Weights35 GB
KV cache32.768 GB
Max context62,000

At 32,768 tokens, the model load uses 71.9% of usable memory. KV cache is 43.8% of the model load.

Planning estimate only. This does not benchmark the Framework Desktop, verify privacy, model quality, runtime allocation, or actual KV geometry.

FIT · 29.232 GB HEADROOM · 62,000 TOKEN LIMIT

Trace entered weight, cache and overhead memory budgets

Read the explanation

This saved planner subtracts the twenty-four gigabyte reserve from one hundred twenty-eight to give one hundred four usable gigabytes. Its default seventy billion parameters at four bits require thirty-five gigabytes of weights. Twenty percent weight overhead adds seven. The entered cache rate and thirty-two thousand seven hundred sixty-eight tokens add thirty-two point seven six eight. Total load is seventy-four point seven six eight, leaving twenty-nine point two three two. Both bars use four pixels per decimal gigabyte. These are entered planning assumptions, not measured hardware allocations. Weight storage equals parameter billions times bits divided by eight. Seventy times four divided by eight is thirty-five gigabytes. The runtime allowance is twenty percent of those weights, or seven gigabytes. Both bars use twelve pixels per decimal gigabyte. Runtime overhead here applies to weights alone, while the cache is a separate added term. The chosen precision does not establish numerical quality or actual deployment compatibility. The pressure button changes context to eighty thousand tokens. At the entered one megabyte per token, cache rises from thirty-two point seven six eight to eighty decimal gigabytes. Both bars use six pixels per gigabyte. Weights and runtime remain forty-two combined, so total load becomes one hundred twenty-two and exceeds one hundred four usable by eighteen. The calculated maximum is sixty-two thousand tokens, obtained by flooring remaining gigabytes times one thousand divided by the entered cache rate. This linear model does not measure real KV allocation.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.