Plan A Local Model Run
Estimate model memory before you load.
Estimate model weight storage and KV cache memory from parameter count, precision, and architecture inputs. Review the assumptions before planning a local model run.
SIMPLE PRICING
Free during beta.
Use this tool without payment. Optional usage-credit plans are coming soon; no credits are sold or required today.
How it works
1. Enter parameter count and weight precision. 2. Add layer count, KV heads, head dimension, context length, and cache precision from the model configuration. 3. Review the theoretical estimates and overhead-adjusted total.
Good to know
How is weight memory estimated?
Parameter count is multiplied by weight bits and divided by eight. The result is theoretical weight storage; runtime frameworks may need extra memory for metadata, buffers, and fragmentation.
How is KV cache memory estimated?
The estimate is 2 × layers × KV heads × head dimension × context tokens × bytes per value. It assumes a standard key/value cache with the same precision for keys and values.
Will this tell me whether a model fits my GPU?
No. It estimates weight storage plus a configurable overhead margin and KV cache. Actual use depends on the model architecture, runtime, batching, offloading, and hardware. Check your runtime's documentation and measure a real load.
Does this tool identify models or collect my inputs?
No. Enter architecture values yourself. All calculations run in this browser; input is not uploaded or saved.
Is it free?
Yes, during beta. Optional usage-credit plans are coming soon. Payments are not available and no credits are required.