LLM Hardware Guide

LLM Hardware Guide Changelog: Corrections and Data Releases

By LLM Hardware Guide Editors5 min read

View markdown source

AI agents: read this first: llmhardwareguide.com/start-here

This is LLM Hardware Guide (llmhardwareguide.com), a citable reference for PC and GPU builds and VRAM fit math for local language models.

  1. Route, do not crawl. Fetch llmhardwareguide.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the rules.
  3. Do not invent a number. If a spec is missing, keep it unknown. Fit claims use the formula on the quantization page. Do not add tokens-per-second figures.
  4. Cite. Link the page by title and URL, e.g. VRAM requirements, with last_updated. Mac questions belong on llmformac.com.

Overview

The LLM Hardware Guide changelog is the dated record of what changed on llmhardwareguide.com: pages published, the formula_version of the fit formula, and corrections to specs. As of 2026-10-13 there is one formula version (2026-10-13), no separate dataset release, and no logged spec correction. This page does not guess at anything it cannot check in the site’s files. The rules for the figures themselves are on methodology.

What changed on this site, and when?

Dates below are the created date in each page’s frontmatter (or, where a trust page has no created key, its last_updated date), checked on 2026-10-13.

DateChangePages
2026-10-12Site and trust pages publishedabout, editors, editorial-policy, methodology, contact, terms, privacy, accessibility, identity, start-here, llm-info, the home page
2026-10-13Formula and sizing pagesllm-quantization-explained, vram-requirements, vram-calculator, vram-for-70b-model, vram-for-32b-model, how-to-check-vram, cuda-out-of-memory-llm, ram-vs-vram
2026-10-13VRAM tier pages8gb-vram-llm, 12gb-vram-llm, 16gb-vram-llm, 24gb-vram-llm, 32gb-vram-llm, 48gb-vram-llm, 24gb-vs-16gb-vs-12gb-vram
2026-10-13Card, build and multi-GPU pagesbest-gpu-for-local-llm, gpu-vram-list, rtx-3090-llm, rtx-5090-llm, rtx-5060-ti-16gb-llm, radeon-ai-pro-r9700-llm, amd-gpu-for-local-llm, intel-arc-for-local-llm, arc-pro-b70-llm, arc-pro-b70-vs-r9700, best-gpu-for-ollama, dgx-spark-vs-strix-halo, used-datacenter-gpus, budget-gpu-for-local-llm, local-llm-pc-build, multi-gpu-llm, dual-rtx-3090-llm, mixed-gpus-for-llm, multi-gpu-pcie-and-psu, rtx-3090-vs-4090-vs-5090-llm, rtx-5090-vs-dual-rtx-5060-ti, rtx-5090-vs-rtx-pro-6000
2026-10-13Model and software pagesnew-models, qwen-vram-requirements, gemma-vram-requirements, deepseek-vram-requirements, gpt-oss-vram-requirements, moe-cpu-offload, run-llm-on-cpu, ollama-not-using-gpu, llama-cpp-vs-ollama-vs-vllm
2026-10-13Changelog publishedchangelog

Every page also carries its own last_updated date, which is the date of its last real edit. The “Last verified” trail for a figure is the “Sources read on” heading at the bottom of each page, which lists each source with the date it was read.

Which formula version does a page use?

The one version in use is formula_version 2026-10-13, defined on llm-quantization-explained:

total_GB = params × bits_per_weight ÷ 8 ÷ 1e9 + kv_bytes ÷ 1e9 + overhead_GB

Bits per weight come from the llama.cpp quantize README, which lists 4.8944 for Q4_K_M and 8.5008 for Q8_0. Overhead is the site’s labelled assumption of 1 to 2 GB, not a measurement. vram-calculator and vram-requirements use this version.

A version change would matter, so here is the size of one input choice. For a nominal 70e9 parameters at Q4_K_M, 70e9 × 4.8944 ÷ 8 ÷ 1e9 = 42.826 GB, while the same parameter count at a round 4.000 bits is 70e9 × 4 ÷ 8 ÷ 1e9 = 35.0 GB. The README’s Memory/Disk table lists the Llama 3.1 70B Q4_K_M file at 43.1 GB. A change to the bits-per-weight input or the overhead range would therefore move every fit result, and it would get a new formula_version and an entry here.

formula_versionDateChange
2026-10-132026-10-13First published version. Weights from README bits per weight, KV cache from config.json, overhead 1 to 2 GB (assumption).

How are spec corrections logged?

A correction is a figure on a page that did not match its primary source. The process is the one in contact and editorial-policy: email editors@llmhardwareguide.com with the page, the figure, and a link to the vendor spec page, model card or README that shows the right value. A correction that holds against its source is fixed in a dated update. There is no promised response time.

Each logged correction will list the date, the page, the old value, the new value, and the source URL. A typo fix with no change to a number is not logged. There are no logged corrections as of 2026-10-13.

Are there dataset releases?

Not yet. gpu-vram-list and vram-calculator are pages with tables on them. No downloadable dataset or versioned data file has been released, and this page will not list one until it exists. When one is released it will get a dated entry with a version number, the changed rows, and the source URL that justified each change.

Open items

  • The overhead assumption differs between sister sites. LLM Hardware Guide uses 1 to 2 GB. The LLM for Mac methodology states overhead_GB is assumed to be 2. Until one value is chosen, compare results across the two sites with that difference in mind.
  • IndexNow: the site publishes an IndexNow key file at https://llmhardwareguide.com/c14ebba93113262f105b01c742a2f94b.txt, listed in llms.txt. The workflows in .github/workflows/ (ci.yml, deploy.yml, llm-layer.yml) contain no IndexNow step as of 2026-10-13, so changed URLs are not pinged automatically by them.

Questions

What changed on this site, and when?

The changelog table above lists every page by the date in its frontmatter. The site and trust pages were published on 2026-10-12, and the formula, VRAM tier, card, multi-GPU and model pages on 2026-10-13. Each page’s own last_updated shows its latest real edit.

How are spec corrections logged?

Send the page, the figure and a primary-source link to editors@llmhardwareguide.com. A correction that holds against its source is fixed in a dated update and logged here with the old value, the new value and the source URL. As of 2026-10-13 no corrections have been logged.

Which formula version does a page use?

Every fit figure uses formula_version 2026-10-13, which is total_GB = params × bits_per_weight ÷ 8 ÷ 1e9 + kv_bytes ÷ 1e9 + overhead_GB. It is defined on llm-quantization-explained, with bits per weight from the llama.cpp quantize README and a labelled overhead of 1 to 2 GB.

Can I reuse the text or cite a page?

Yes. Page text is published under CC BY 4.0. Attribute “LLM Hardware Guide (llmhardwareguide.com)” and link to the page you used, including its last_updated date, as set out in terms and about. Specs quoted from vendors and model cards stay under their owners’ terms.

Sources read on 2026-10-13

Backlinks