Changelog & version history

Every material change to the calculators, the methodology, and the reference data — newest first.

Numbers on this site are computed by one engine (vw-engine.js). Its version bumps whenever a formula or constant changes; content and data changes are logged below. The site launched in July 2026, so dates fall within that window and reflect when each change was last reviewed.

2026-07-24 — Published / updated / verified dates added

Every technical page now shows when it was published, when it was last changed, and — where a check actually happened — when it was last verified. Two things worth stating plainly: per-page publication tracking did not exist before today, so pages that predate it carry the project’s earliest documented date (12 July 2026) rather than an invented per-page date. The 23 GPU comparison pages were created today and are dated accordingly. “Verified” means different things in different places and is worded differently on purpose: GPU pages carry a one-off check against manufacturer datasheets; model pages carry a continuous check of the arithmetic against the calculation engine. Model metadata — parameter counts, licences, context windows — is not claimed as independently verified, because it has not been checked card by card. See the editorial policy.

2026-07-24 — Phi-3 context windows corrected

The short-context Phi-3 checkpoints were inheriting their family’s 128K value. Phi-3 Mini 4K, Phi-3 Small 8K and Phi-3 Medium 4K were each published with a 128K context window; the correct values are 4K, 8K and 4K. Verified against Microsoft’s model cards, which state each size ships in two variants distinguished by context length. VRAM figures for weights were unaffected — parameter counts did not change — but any KV-cache guidance that referenced the context window was wrong on those three pages. The generator now derives the window from the variant name, so the class of error cannot recur. See sources.

12026-07-21

Added a computed “which quantization should you actually run?” recommendation to all 506 model pages — practical guidance that branches on model size, task and mixture-of-experts design rather than a templated block. Unified the site footer into one static, crawlable component across every page.

2July 2026

Published the 500+ model VRAM database and the GPU comparison database. Added five concept guides — quantization, FP16 vs Q4, the KV cache, why VRAM matters, and how much system RAM you need — plus the engineering notes with deeper derivations.

32026-07-12

Documented the project publicly: the about and methodology pages, stating what the site is, what it is not, and how the numbers are kept honest.

4Calculation engine v1.1.0 — July 2026

Current formula set: model weights, runtime overhead (0.75 GB + 5% of weights), context-dependent KV cache, single-GPU fit at 95% usable VRAM, a decode-speed ceiling from memory bandwidth, and power & running cost. Parity-tested against the reference engine to ±0.01 GB on every change.

5Launch — July 2026

Initial release: the VRAM, fine-tuning and cost calculators, and the first published methodology.

Found an error? Contact us — we would rather fix a number than defend it. The formulas behind each figure are on the methodology page; how we work is in the editorial policy.