Home › Changelog

Changelog

Every change to this site gets a dated line here — including the ones that make us look worse. Why: our whole claim is "every figure has a source and a check date", which is worthless if you cannot see when a number moved. Last entry: 2026-10-23.

2026-10-23

• Added three runtime compares — Ollama vs llama.cpp, Ollama vs LM Studio, Ollama vs vLLM — built only from official docs pages we fetched on 2026-10-23 (READMEs and docs links listed on each page). Sitemap went from 44 to 47 URLs. We publish no tokens/s figure anywhere: nothing comparable was verifiable, so speed rows are absent rather than guessed, and cells we could not cite read "no public data (we checked, 2026-10-23)".
• Added a fourth runtime compare — vLLM vs llama.cpp — so the four engines now sit in one closed link set; sitemap went from 47 to 48 URLs. Same rule as the other three pages: only official docs fetched 2026-10-23, no tokens/s figure anywhere, and every cell we could not cite printed as a gap.
• Added an “On an Apple Silicon Mac” row to all four runtime pages, quoting one Apple-specific sentence per engine from the same four official pages fetched 2026-10-23: Ollama GPU docs (“Metal (Apple GPUs)” — GPU acceleration on Apple devices via the Metal API”), llama.cpp README (“Apple silicon is a first-class citizen”), LM Studio docs (llama.cpp or Apple’s MLX on Apple Silicon Macs), and vLLM’s install page (Apple Silicon delegated to vLLM-Metal, “a community-maintained hardware plugin…”). Reason: people pick a runtime per machine, and on a Mac system memory doubles as GPU memory, so the discrete-card rows above do not answer their question.
• Added four Mac pages: Mac mini M4 16GB, Mac mini M4 24GB, Mac mini M4 Pro 48GB and MacBook Pro M4 16GB — each one lists which sourced models fit that machine and which do not. Sitemap went from 48 to 52 URLs. Memory configurations are quoted from Apple’s own tech-spec pages fetched 2026-10-23; the GPU-usable fraction Apple does not publish, so it is printed as “no public data” and our usable-memory heuristic is labelled as ours on the page. No tokens/s figure, same as everywhere else on this site.
• Added the runtime comparison table — one grid, four engines, 8 rows — because the autocomplete sets we logged on 2026-10-23 include searches naming three engines at once (e.g. “llama.cpp vs ollama vs vllm”, 4 suggestions), which no pairwise page answers on its own; the four pairwise pages also gained a parent page above them. Sitemap went from 52 to 53 URLs. No new sources were fetched: each cell restates one statement already taken from official docs on 2026-10-23 and carries the same superscript link, while the pairwise pages still hold the complete sentences. 6 of the 32 cells read “no public data”, printed as gaps, and the grid publishes no tokens/s figure — same rule as the rest of the site.
• Added the Mac pages hub — one row per machine, with Apple’s own memory line, the usable budget we apply and how many dataset models fit — because the four machine pages had no entry point above themselves. The home page and the footer now link it, and link the runtime table too: an audit after our last deploy counted nine pages added that night with no link at all from the home page (measured 2026-10-23), which is a link-graph defect, not a cosmetic one — a page nobody links to is a page nobody reaches. Sitemap went from 53 to 54 URLs. No new sources were fetched: every cell restates an Apple spec line or a fit count already published on a machine page on 2026-10-23. One machine is named as missing on purpose: we still have no quotable Apple spec sheet for the MacBook Air (M4), so it has no page — no spec sheet, no page, same rule as everywhere else here.
• Corrected the <link rel=canonical> target on every page but the home page. Each page pointed at its own /index.html file, which the host answers with a 307 redirect to the clean URL (checked with curl on 2026-10-23), while sitemap.xml listed the clean URL — two different answers to “which is the canonical page”. Canonicals now match the sitemap URL exactly, and the build gate compares the two so they cannot drift apart again.

2026-10-22

• Added 6 pages: DeepSeek V4.1 Flash, Kimi K3 q4_K_M, Kimi K3 on 24GB, and three cloud-provider compares (RunPod vs Vast.ai, RunPod vs Lambda, Vast.ai vs Lambda). Sitemap went from 37 to 43 URLs.
• Added a visible cloud-rate check date block on the home page and the compare pages. Reason: the third-party rate table we cite states its own check date of 2026-09-29, and our re-fetch on 2026-10-22 showed the values unchanged — so those "2026" rates are 3+ weeks old and now say so on the page instead of implying freshness.
• Published no figure at all for Vast.ai hourly GPU rates. Its pricing is per-listing and its pricing page is JS-rendered, so we could not verify a citable number as of 2026-10-22; the affected cells read "no citable rate — no public data" rather than a plausible estimate.
• Corrected a truthfulness bug we had shipped: a cross-check footnote documenting third-party figures for one specific model family was being printed in the footer note of every model page, including unrelated models. It now appears only on the two pages that family covers.

2026-10-21

• This site went live at llmbudgetlab.com on Cloudflare static assets: bare domain and www both bound, always-HTTPS on, and a www → bare-domain redirect (verified by our owner, 2026-10-21).
• Interface v3 shipped: system font stack (no webfont download), light theme by default with a remembered dark option, tabular figures in tables, keyboard focus rings. No trackers, no cookies, no account.
• Published the money-facing plumbing that this changelog depends on: a pre-deploy check that refuses to publish any price figure that is not registered with a named source on the method page, plus clearly labelled outbound provider links and a sponsor page that forbids unlabelled placement.

2026-10-20

• Data baseline for every VRAM figure on this site: the modelvram.com open dataset (CC BY 4.0), retrieved 2026-10-20, plus the rental-rate table whose own stated check date is 2026-09-29. Our extrapolation of KV-cache growth with context length is our own, published on the method page — and we state there that it mis-models MLA / sparse attention.
• Added the regression harness that runs before every deploy: it recomputes 39,744 model × hardware × quantisation × context combinations through the same calculator code the browser uses and diffs them against the shipped math, so a page can't quietly disagree with the table next to it.

What we log and what we don't

We log anything that changes a number a reader might act on (a rate, a GiB figure, a source, a check date, a rule like our usable-VRAM rule), plus every correction we owe you. We don't log typos or CSS tweaks — if it cannot change someone's buy/rent/API decision, it isn't a changelog line.

Corrections

Spotted a wrong number or a stale date? Mail hello@llmbudgetlab.com with the page and the figure. If you are right it gets fixed, credited here, and the source line changes — that is the deal this page exists to keep.