Controlled public beta
Known limits
InferGrade helps technical users choose and compare local LLM setups with evidence. It does not certify that a model is safe, reliable, legally suitable, or ready for operational use.
Supported path
- Apple Silicon with the Desktop Runner is the reference beta path.
- Managed llama.cpp Metal runtime setup is supported after explicit user action.
- Hub queued runs require a paired, listening Runner with a supported contract. The latest release is recommended; the tagged OpenAI Build Week v0.3.39 submission remains compatible during judging but may not support newer models or runtimes.
- New runs stay private while executing and publish accepted results by default when complete. Owners can opt out before queueing or unpublish later; existing private results are unchanged.
Preview and deferred paths
- Windows with NVIDIA CUDA is a technical beta target, not a controlled-public-beta feature yet.
- Linux CUDA remains a best-effort CLI path until there is a separate support plan.
- ROCm, Vulkan, non-Mac CPU-only paths, and build-from-source runtime setup are experimental unless a result says otherwise.
- Cloud execution is deferred unless an internal run is explicitly marked experimental.
Evidence shapes
- Demo, sample, thin, failed, and informational evidence stay visibly distinct from decision-grade evidence.
- Result pages describe what a run proves, what it does not prove, and what benchmark would improve the answer.
- Scores are scoped to the use case, benchmark lane, hardware, runtime, and artifact that produced the evidence.
- InferGrade does not publish raw private outputs, runner tokens, pairing codes, signed URLs, or private artifacts.
Before relying on an answer
- Check whether the Result page is exact, nearby, sample, or informational evidence.
- Compare only within the stated use case, hardware, runtime, and benchmark scope.
- Run the suggested follow-up benchmark when the answer asks for confirmation.