About LLMHardware.io
By LLM Hardware team · Updated
What We Do
LLMHardware.io is an independent reference site that helps people choose the right hardware for running open-source large language models locally — on their own machine, without cloud APIs.
Kept guides and the VRAM calculator size a model from its parameter count: Q4, Q8, and an FP16 sketch, plus context. There is no page per Hugging Face model. Recommendations use published VRAM and bandwidth, and the site formula.
Our Methodology
Hardware recommendations are generated by a rules-based VRAM calculator using the formula:
VRAM (GB) = parameters (B) × bytes_per_param × 1.2 overhead
Where bytes_per_param is 2.0 for FP16, 1.0 for Q8, 0.625 for Q5_K_M, and 0.5 for Q4_K_M. An additional 2 GB is reserved for OS and inference engine overhead.
Hardware specifications — VRAM capacity, memory bandwidth, TDP — are sourced from manufacturer spec sheets. GPU pages show the manufacturer list price with a source and a check date. They do not show an Amazon price. Each product link says "Check current price on Amazon" and goes to Amazon for the live price.
The site’s model list was imported from the Hugging Face Hub API. Its automated refresh has been off since 2 May 2026, so that list is no longer updated automatically. Parameter counts in that list were parsed from each model’s Hub ID and tags (for example the “8B” in Qwen3-8B), so they are estimates, and some models have none. Some guides link individual model cards, and hardware data is updated in the weekly maintenance run.
Who runs this site
LLMHardware.io is written and maintained by the LLM Hardware team. The team researches published specs, vendor spec sheets and model cards, and runs the numbers through the site's calculator and VRAM formula (params × bytes × 1.2 + 2). Drafting is AI-assisted, with an edit pass before publishing. This site does not measure hardware.
Reach the team via the contact page.
No vendor sponsorships, no paid placements. Hardware recommendations follow published VRAM, memory bandwidth, and the manufacturer list price. Affiliate commissions on Amazon links do not change which hardware appears in which tier; see the methodology page for the full ranking process.
Where else to find us
Our only other official channel is the companion blog, Run Open Source LLM, which links back here. Author page: LLM Hardware team. Any other account using our name is not ours.
What We Cover
- VRAM-fit tiers for popular open-weight models, via the VRAM calculator and fit checker
- Consumer GPU comparisons based on published specs and the VRAM formula (NVIDIA, AMD)
- Apple Silicon for LLM inference (M3, M4, and M5, including unified-memory configs that fit a given quant)
- Mini-PC and budget builds for local AI
- Quantization guides and VRAM optimization techniques
- Setup guides for Ollama, llama.cpp, LM Studio, and related tools
Affiliate Disclosure
LLMHardware.io participates in the Amazon Associates program. When you purchase through our links, we may earn a commission at no additional cost to you. This helps us maintain the site and keep content free.
Affiliate relationships do not influence our recommendations. We link to the hardware our methodology identifies as the best fit for each use case — not to the products with the highest commission rates. See our full disclosure page for details.
Data Accuracy
GPU pages show the manufacturer list price with a cited source and the date it was checked. They do not show an Amazon price. Every Amazon link is there so you can check the live price and availability. Confirm that price on the retailer before purchasing.
Model parameter counts and VRAM requirements are calculated estimates. Actual memory usage may vary depending on inference engine, context length, batch size, and other factors. Our figures are a practical starting point, not a guarantee.
Contact
For corrections, partnership inquiries, or general questions, visit our contact page.