OUR METHODOLOGY
A recommendation is only
as useful as its evidence.
We keep purchase availability, model compatibility, and measured performance separate. Here is what the first release can—and cannot—tell you.
Availability has a place and a time.
The initial retailer integration is Best Buy store pickup in the United States. Other countries listed in the market selector are planned and cannot be searched yet. Locale and purchase currency are separate; this interface is English. An offer must match your ZIP, a specific store, and the exact hardware configuration. A national product listing or an online availability flag is not proof of delivery to your address.
Only current, successfully verified snapshots are eligible. Expired snapshots, disabled retailers, invalid prices, and failed latest checks are excluded. We recheck availability before sending you to the retailer. Stock and prices can still change at checkout.
Your workload changes the answer.
A model search uses public Hugging Face metadata. You choose the exact repository; we do not silently pick a similarly named model. Maximum total tokens includes both the prompt and generated output for each sequence. Concurrent requests means the maximum number of simultaneously resident sequences.
Memory estimates account for the model weights and workload-dependent memory. Original precision and verified quantized artifacts are distinct. Missing model facts or unverified runtime support result in an unknown assessment, not an execution claim.
Compatibility is not a speed benchmark.
Positive recommendations require supported execution evidence. Unknown and unsupported assessments are excluded from model recommendations and explained. Browsing without a model can show verified offers without claiming they can run a particular model.
No throughput or performance-per-dollar number is invented. Within an execution tier, offers without matching measured performance use price plus shipping, then offer ID, as the deterministic fallback order. Each offer shows its currency and whether tax is included. Current US pickup totals are before tax. We do not convert currencies.
A deliberately narrow first release.
The current scope is public LLM metadata, conservative model assessment, and verified single-PC store-pickup offers. There is no live PC inventory in the current early-access deployment. A retailer API connection and collected, current snapshots must be available before any products can appear. API configuration does not imply affiliate approval. Affiliate links and commission-based ranking are not configured in this release.
Multi-GPU recommendations, CPU offloading, nationwide delivery claims, and broad retailer coverage are not supported in this release. Hardware-specific performance calibration and expanded execution support are planned; they are not presented as completed capabilities.