Methodology & data sources
How we source model specifications, calculate memory estimates, and report benchmark results.
Published.
Specifications from the model publisher. We link to the original model card and distinguish total parameters from active parameters.
SOURCE-LINKEDCalculated.
Weight storage is parameters × bits ÷ 8, in decimal GB. It excludes KV cache, runtime allocations, and quantization metadata.
AN ESTIMATEMeasured.
Performance requires a reproducible experiment. No BenchGrid performance runs have been published in this first edition.
NOT YET AVAILABLEWhat a future benchmark must include.
Model and tokenizer revisions; exact GPU and instance configuration; runtime, CUDA and driver versions; precision; parallelism; context and output lengths; arrival rate and concurrency; warmup; cache policy; errors; repeat variability; and raw results.
We will report latency and throughput separately, with the workload and measurement boundary visible. A price-derived cost will be labeled as calculated, with its region, pricing basis, and date.
Reference: vLLM benchmark documentationA weight estimate is not a GPU recommendation.
The memory explorer illustrates how bit width changes weight storage. It does not establish checkpoint availability, output quality, runtime compatibility, or a successful fit. Nominal sizes are labeled with ~. MoE calculations use total parameters for a fully resident model; offloaded deployments need separate analysis.
Independence & commercial links.
The provider links in this first edition are ordinary links. BenchGrid has not connected an affiliate account, receives no tracked commissions from these links, and has no sponsored placements.
If affiliate links or sponsored content are introduced, they will be labeled. Payment will not change measurements or turn an untested configuration into a recommendation. Results from one cloud will not be represented as measurements from another.
A living field guide.
The homepage features six editor-selected Trending models. These picks are not a live popularity ranking. The full directory contains 20 current candidates and six established baselines. New profiles keep memory estimates pending until their checkpoint scope is reconciled. Sources were reviewed on September 22, 2026.
Explore all models