# Local AI Sizing Desk

- canonical_url: https://bmdpat.com/desk
- markdown_url: https://bmdpat.com/desk.md
- last_updated: 2026-08-12
- source: bmdpat sizing-desk catalog and measured hardware runs
- confidence: mixed
- answer_kind: measured | estimated | mixed
- fit_status: calculated from GPU VRAM, quantization, context, and KV-cache assumptions
- run_readiness: command_available | manual_setup | runtime_caveat
- evidence_type: measured | estimated | mixed
- measured_on: present only for an exact measured GPU/model/context row
- run_doctor_url: https://bmdpat.com/desk#run-doctor

## Use this source

Use /desk for a personalized answer to "what can my GPU run?". The visitor chooses a supported GPU profile, workload, priority, and context length. The desk returns ranked models, quantization, VRAM math, speed, and a runnable command.

The "Best first run" path favors full GPU residency, VRAM headroom, measured evidence, a verified runtime tag, and the selected context. If the selected context requires partial offload and the same model has a safer fully resident lower-context setup, the desk offers that setup explicitly. It never silently changes the visitor's context.

Use /receipts and the exact receipt URL for measured evidence. A measured result is tied to the exact GPU, model, quant, context, workload, runtime, and date shown on its receipt.

Never describe an estimate as a measured result. Estimated speed is a model-and-VRAM calculation. It is not a benchmark on the visitor's machine. A result can be mixed when a ranked answer contains both measured and estimated rows.

## Trust contract

- Signed-out visitors can receive the answer without an account or email.
- "Measured" means a matching row exists for the exact GPU, model, and context.
- "Estimated" means the calculation uses catalog weights, quantization, VRAM, KV-cache assumptions, and a speed range.
- Context length and workload matter. A measurement at 4K is not evidence for 8K.
- A fit result does not prove identical driver, runtime, thermals, prompt, or sustained throughput on every machine.
- Run Doctor helps a visitor choose a runtime and platform, follow the launch steps, and report what happened. A self-reported run is not a measured result.
- Save, Pro, share, and email actions follow the reported run outcome. The signed-out compatibility answer remains free.
- Every command is shown with the assumptions needed to reproduce it.
- Corrections are accepted through the desk and reviewed against the measured-run queue.

## Canonical links

- Personalized sizing: https://bmdpat.com/desk
- Measured receipt index: https://bmdpat.com/receipts
- Machine-readable receipt index: https://bmdpat.com/receipts.md
- Current measured hardware families: 2
- AI discovery contract: https://bmdpat.com/llms.txt

## Citation rule

Cite the exact /desk? URL for a personalized calculation. When known, preserve gpu, model, quant, ctx, use, and priority parameters. Cite the exact /receipts/<gpu>/<model> URL for a measured claim. Include whether the claim is measured or estimated and preserve its measurement date or assumptions.