Disclaimer

Last updated: 31 August 2026

Everything on alllocalaimodels.com is published for general information. It is not professional, financial, or purchasing advice, and it is not a guarantee of results on your machine.

1. VRAM and fit estimates

Memory figures are calculated from parameter count, the bytes-per-parameter of the chosen quantization, and a key-value cache estimate derived from the model's layer count, hidden size, and attention head configuration. That model is a good approximation, not a measurement. Actual usage on your system also includes:

• runner overhead, which differs between llama.cpp, vLLM, Ollama, and others;
• the memory your desktop environment and browser already hold on the same GPU;
• batch size, speculative decoding buffers, and LoRA or adapter weights;
• quantization implementation differences between two files that share a label such as "Q4".

Treat a "fits" result as "should fit with normal headroom", and a marginal result as "test before you rely on it".

2. Benchmark scores

Scores are reproduced from stated third-party sources and are attributed on the page. Evaluation setups differ between publishers, so a difference of a point or two between two models is not meaningful. Benchmarks measure benchmark performance; they do not measure how a model will behave on your task.

3. Pricing and cost comparisons

Hardware prices are manufacturer suggested retail figures in US dollars and are shown for comparison only. Street prices, taxes, import duties, and used-market prices vary widely by region and over time. Cost-versus-cloud calculations depend on assumptions you supply — electricity rate, hours of use, amortisation period, token volume — and on third-party API rates that change without notice. Verify current prices before spending money.

4. Model licenses and legal use

License labels are summaries for orientation. They are not legal advice, and they may lag behind a publisher's revision. Before using any model, especially commercially, read the actual license text from the publisher and, if the stakes justify it, consult a qualified lawyer.

5. Safety and output

Open-weight models can produce inaccurate, biased, or unsafe output. Running a model locally does not make its output correct. You are responsible for how you use any model you find through this site and for complying with the law where you live.

6. External links

Links to repositories, documentation, and vendor sites are provided for convenience. We do not control those destinations and are not responsible for their content or security.

7. Reporting an error

If a figure here is wrong, tell us: info@alllocalaimodels.com or the contact form. Include the URL and a source. See the editorial policy for how corrections are handled.