Corrections and Updates

Last updated: · Maintained by the LLM Hardware team

Why this page exists

Hardware specs change. Prices move. A published benchmark can turn out to have used the wrong inference flag. This page is the public log of substantive corrections made to guides on LLMHardware.io after publication. The log starts with the November 2026 fit corrections; earlier fixes were made in place and are not all listed. The goal is that anyone who read an old version of a guide can come here, see what changed, and decide whether they need to revisit a purchase decision.

Trivial fixes (typos, broken non-load-bearing links, formatting) are not logged here. They are quietly fixed in place. Anything that changes a recommendation, a price, a benchmark number, a compatibility claim, or a hardware spec is logged below with the date and what changed.

How to report an error

Spotted something wrong? Use the contact page with:

  • The page URL
  • The exact claim that is wrong
  • What the correct value or claim should be, with a source if you have one (vendor spec sheet, benchmark thread, your own measurement)

Confirmed corrections that change a recommendation or a fit claim are logged below with the date and the pages changed.

Correction log

2026-11-09 · /guides/mistral-hardware-requirements

Was: The buying advice said Mistral 7B at Q8 (about 8 GB) fits an 8 GB RTX 4060, suggested that card for Nemo 12B at Q4, put Mistral Small 22B at Q8 on a 24 GB RX 7900 XTX, gave a 16 GB card about 3 GB of room for Small 22B at Q4, and gave a 24 GB Mac about 11 GB of headroom for Small 22B at Q4.

Now: 7B at Q8 is 10.4 GB and Nemo 12B at Q4 is 9.2 GB, so both need a 12 GB card. Small 22B at Q8 is 28.4 GB and needs a 32 GB card. Small 22B at Q4 (15.2 GB) leaves about 0.8 GB on a 16 GB card, and about 8.8 GB on a 24 GB Mac before the macOS reserve (about 4.8 GB after it). Source: the page's own table and the site VRAM formula.

Found by: internal review

2026-11-09 · /guides/qwen3-hardware-requirements

Was: The FAQ said Qwen3-235B-A22B requires 4 or more 24 GB GPUs. The VRAM tables were labelled as weights only, and the Qwen3-4B row listed a 4 GB minimum beside a 4.4 GB estimate.

Now: Qwen3-235B-A22B is about 143 GB at Q4, more than four 24 GB GPUs (96 GB) hold. The tables are labelled as site-formula estimates that include a 2 GB allowance, and the Qwen3-4B minimum is an 8 GB GPU or CPU. Source: the site VRAM formula.

Found by: internal review

2026-11-09 · /guides/best-gpu-for-llms

Was: A 64 GB Mac ran 70B at Q4_K_M with 27 GB of headroom, and running 70B without multiple GPUs required Apple Silicon. The RTX 4070 12 GB card listed about 3 GB of headroom for 13B at Q4.

Now: 70B at Q4 is about 44 GB by the site formula, which leaves about 20 GB on a 64 GB Mac before macOS and about 16 GB after a 4 GB reserve. A Mac with 64 GB or more or the NVIDIA DGX Spark (128 GB) can hold it on one device. 13B at Q4 is 9.8 GB, which leaves about 2.2 GB on 12 GB. Source: the site VRAM formula and the NVIDIA DGX Spark spec page.

Found by: internal review

2026-11-09 · Home page

Was: The headline said "Find Hardware That Runs Any LLM Locally", and the privacy answer said local LLMs are 100% private and your data never leaves your computer.

Now: The headline is "Find the Hardware Your Local LLM Needs". The privacy answer separates local models from cloud features and connected tools, which can send data off your computer. Source: the Ollama FAQ.

Found by: internal review

2026-11-06 · Model-fit and bandwidth corrections

What changed: Remaining model-fit and bandwidth claims were corrected. Pages changed in that update: /guides/apple-silicon-for-llms, /guides/best-llm-for-data-analysis, /guides/best-llm-for-writing-locally, /guides/gemma3-hardware-requirements, /guides/gguf-vs-gptq-guide, /guides/llm-quantization-guide, /guides/lm-studio-vs-ollama, /guides/mistral-hardware-requirements, /vram-calculator, and the hardware data.

The earlier wording is not reproduced here.

2026-11-06 · Guide fit claims and recommendation method

What changed: Guide fit claims and the recommendation method were corrected, and navigation was repaired. Pages changed in that update: /guides/apple-silicon-for-llms, /guides/beginners-guide-to-local-ai, /guides/best-budget-gpu-for-llms, /guides/best-gpu-for-llms, /guides/best-llm-for-coding, /guides/best-llm-for-data-analysis, /guides/best-llm-for-writing-locally, /guides/best-llms-to-run-locally, /guides/deepseek-hardware-requirements, /guides/gemma3-hardware-requirements, /guides/gguf-vs-gptq-guide, /guides/how-to-run-llms-locally, /guides/llama-3.1-hardware-requirements, /guides/llm-quantization-guide, /guides/llm-system-requirements, /guides/lm-studio-vs-ollama, /guides/mac-vs-pc-for-llms, /guides/mistral-hardware-requirements, /guides/phi4-hardware-requirements, /guides/qwen3-8-flash-next-hardware-requirements, /guides/qwen3-hardware-requirements, /guides/rtx-3090-llm-guide, /guides/rtx-4090-llm-guide, /guides/strix-halo-mini-pcs, /guides/what-can-i-run, /hardware and the hardware product pages, /methodology, /editorial-policy, and the hardware data.

The earlier wording is not reproduced here.

2026-11-05 · Model fit and quantization guidance

What changed: Model fit and quantization guidance was corrected. Pages changed in that update: /guides/how-to-run-llms-locally, the home page, and /vram-calculator.

The earlier wording is not reproduced here.