Show the math. Keep the uncertainty.

Memory fit

Required bytes = recorded weight bytes + KV-cache bytes + runtime overhead.
KV bytes = 2 × layers × KV heads × head dimension × context × bytes per element.
Usable bytes = installed GiB × 2^30 × usable fraction.

The default is one sequence at up to 4,096 tokens, 16-bit KV elements, and 1 GiB runtime overhead. Most unified-memory entries reserve 25% for the system; the hardware record states its exact fraction. Dedicated GPUs normally use a 90% memory budget. This is a configurable planning assumption, not a guaranteed runtime allocation limit.

“Fits” means the estimated requirement occupies at most 90% of the usable budget. “Borderline” means it fits but exceeds that threshold. “Doesn't fit” means the known requirement exceeds the supported budget. “Fits with offload” is conditional and requires specified host memory and bandwidth. Unknown cache geometry, unsupported layouts, vision budgets, or multi-GPU placement remain unknown.

Estimated generation speed

Estimated tokens/s = bandwidth bytes/s ÷ (active parameters × effective bytes per weight) × efficiency.
Effective bytes per weight = recorded file size ÷ total parameter count.

Default efficiency is 0.6. The default displayed range uses 0.4–0.8, a sensitivity range, not a statistical confidence interval. The file-size ratio includes metadata, mixed-precision tensors, and alignment. All MoE experts count toward weight memory; only the active-parameter proxy changes estimated decode traffic.

This model does not predict prefill, batching, simultaneous users, training, image encoding, kernel behavior, thermal throttling, or attention cost at long context. Actual performance can be outside the range. Offload uses a conservative serial GPU-plus-host full active-weight pass; transfer latency and CPU compute are not modeled.

Units and provenance

Memory capacities labeled GB by a maker are interpreted as GiB for the estimator; memory bandwidth is decimal GB/s. The JSON contract retains memoryGB field names, with this unit convention. Hardware field sources distinguish official values, derived bandwidth, peak power, and unknown prices. Power figures are not measured inference consumption.

Model discovery combines a download-ranked GGUF feed with major-lab queries. The catalogue is an organization-source-filtered discovery sample, not a claim to cover every popular model or an exact global top 150. Base models are deduplicated; quantizations use complete pinned file listings. Overrides are versioned separately. Unknowns are preserved.

Commercial terms and affiliate links

Permissive license metadata is labeled subject to the official terms. Custom or missing terms require review; no commercial-use clearance is implied. There are no paid placements in this build. Affiliate links are inactive without an enrolled publisher's tag and verified product links.

Published benchmarks

The table below shows outside measurements for context alongside our estimate where the corresponding model is available. The quantizations and software differ, so no error percentage, calibration, or accuracy conclusion is valid. Web Asset House did not perform these hardware benchmarks.