Directional deployment planning
Fit the model before the benchmark.
Allocate hybrid precision across an MoE model, expose the binding constraint, and leave with a benchmark-ready engineering brief.
Calculating sample
Model and workload
Deployment inputs
Current estimate
Fits with room to test
MXFP8
NVFP4
NF3
Balanced tradeoff
Change an input to inspect the recommendation.
Scenario comparison
Three ways to deploy
Transparent assumptions
Know what is estimated
Weight residency
Uses nominal 8, 4, and 3-bit weights plus a 6% metadata allowance.
Runtime and KV cache
Approximates cache from active parameters, context, and concurrency; runtime adds 4.5% of weights plus 2.5 GB.
Quality and speed
Heuristic format penalties and multipliers support planning only. Validate on your own tasks and kernels.
Durable decision
Save the evidence
From estimate to evidence
Benchmark the constraint, not the whole unknown.
Your deployment brief will update with the current scenario.