Editorial standards

State calculation assumptions first

VRAM is separated into approximate weight memory, calculated KV cache and a planning reserve. Calculations use bytes and display GiB. We disclose factors that can vary with hardware, runtime, batches and other applications.

API estimates disclose the official rate check date, currency, billing scope and cache-hit assumptions. The 5-minute/1-hour comparison models sequential requests with a fixed prefix, not paid API calls or verified bills. RAG estimates disclose equal document lengths, fixed token windows, document boundaries, final chunks, duplicated tokens and document-count rounding; they do not reproduce a specific splitter or bill. Rates are normally checked monthly; dates are never refreshed without checking sources.

Identify models and distribution artifacts

• We check model cards, config.json, official libraries and distribution tags. If access is restricted, alternative evidence and its limits are disclosed.
• Exact GGUF byte sizes and identifiers are recorded. Files from different publishers with the same name and precision are not treated as identical.
• Unavailable artifacts, unmeasured results and unsupported calculations are distinguished. Missing data is not invented or presented as observed fact.

Make measurements reproducible

• Record GPU, CPU, RAM, OS, driver, exact Ollama version and measurement date.
• Check baseline VRAM with other models unloaded, actual input/output tokens, context, F16 KV, Flash Attention, concurrent requests and model placement.
• Observe loading, prompt processing and generation, then subtract baseline memory from the highest sampled usage. Record sampling intervals and repetitions.
• Publish the largest context tested and observed prediction error. Results from one device are not generalized to every environment.

AI assistance and verification responsibility

AI assists research, calculation implementation and editing. Official sources and calculated results are cross-checked and kept separate from hardware measurements.

AI rechecking is not described as independent human review. We do not invent product experience, experts, qualifications or performance figures. Operator details are published on the About page.

Check evidence before correcting errors

Changes to source artifacts or runtime behavior trigger a review of affected calculations and explanations. Source check dates change only when sources are actually checked again.

For corrections, send the page URL, problematic value, selected inputs and public supporting evidence. Secret keys and private files are unnecessary.