Suggesting a New Model or Hardware Configuration
We continually update our database as new frontier open-weights models and GPUs release. When requesting a new model, please include the Hugging Face repo, architecture (Dense vs. MoE), active parameters, and attention geometry.
Reporting a Bug or Discrepancy
If you observe a discrepancy between our estimated tokens/sec throughput and your physical hardware benchmarks, please let us know with your GPU model, driver version, inference engine, and batch configuration.