About All Local AI

All Local AI Models exists to answer one question honestly: will this model actually run on the hardware I own?

Most model listings are copied from a model card. They tell you the parameter count and stop. That leaves you downloading a 40 GB file to find out your 12 GB card cannot hold it once the KV cache is allocated. This directory computes the memory requirement for every quantization, compares it against real GPU specs, and says plainly whether it fits, fits with a short context, or does not fit at all.

What you will find here

• Models — parameter count, active parameters for mixture-of-experts architectures, context window, license, and VRAM per quantization level, plus how to run each one.
• Tools — the inference engines and desktop apps that load those weights: install method, operating system support, supported model formats, and what each one is genuinely good at.
• Hardware — GPU listings with VRAM, memory bandwidth, power draw, and MSRP, with the published models that fit each card.
• Calculators — the same VRAM engine the model pages use, exposed directly so you can test a model, quantization, and context length combination before committing bandwidth to it.

Where the numbers come from

VRAM estimates are computed, not copied. The engine multiplies parameter count by bytes-per-parameter for the chosen quantization, then adds key-value cache memory derived from layer count, hidden size, attention head configuration, and the context length you ask for. Architectural values come from published model configuration files. Where a value is unknown, the page says unknown rather than guessing.

Benchmark scores are only shown when they come from a stated source — an official model card, a maintained leaderboard, or a paper. Scores are attributed on the page. We do not run private benchmarks and present them as objective truth, and we do not average scores from incompatible evaluation setups.

Hardware pricing is manufacturer suggested retail pricing in US dollars. Street prices move constantly and vary by region, so treat MSRP as a reference point for comparison between cards, not as a quote.

How this is different from a model hub

Hugging Face hosts the weights and is the correct place to download them. This site is the layer above that: an opinionated, computed view of what those weights cost in memory, which runner handles them best, and which card you need. Every model page links back to its source repository, because you should verify anything that matters to you.

Independence

All Local AI is independently operated. Listings are not paid placements. No vendor can buy a ranking, a score, or a position in a comparison. The site is funded by advertising, which is kept visually separate from editorial content and never influences which models or tools are covered. If that ever changes, it will be disclosed on this page before it takes effect.

Corrections

Specifications change: models get re-released with different configs, licenses get revised, tools change their install path. If a number here is wrong, we want to fix it. Send the page URL, the value you believe is incorrect, and a source at info@alllocalaimodels.com or through the contact form. Corrections with a verifiable source are prioritised over everything else in the queue.

Our full standards for research, review, and AI assistance are documented in the editorial policy.

Start here

If you know your GPU, run What can I run. If you know the model, use the VRAM calculator. If you are still choosing a card, read the GPU buying guide and then browse hardware.