What Local AI Fit does
Pick your graphics card, Mac or RAM and see which open AI models run on it and how fast. Or pick a model and see what hardware it needs.
Running an open AI model on your own computer comes down to one question: does it fit in memory, and is the memory fast enough? Local AI Fit answers both. Pick the graphics card, Apple chip or system RAM you have and it lists thirteen popular open models, the quantisation that fits, and a speed estimate in tokens per second.
The arithmetic is shown, not hidden. Weights are parameters times bits per weight. The context cache is layers times KV heads times head size times context length. Add runtime buffers and compare with the memory you have. When a model is larger than your graphics card's memory it can spill into system RAM; the tool shows how much of the model sits on the card and how much that slows generation.
Reverse mode is for buying decisions. Choose the model you want to run, the quantisation and the context length, and it shows the memory required plus the graphics cards and Mac configurations that hold it entirely. Prices are not shown because they move too fast; the memory figure is what stays true.
Speeds are ceilings estimated from memory bandwidth, not benchmarks. Treat them as a way to compare options and to tell a usable setup from an unusable one, not as a promise.
Popular uses
Each opens Local AI Fit already set up for that job, with steps and answers for that case.
Free vs Pro
| Feature | Free $0 · no account |
|---|---|
| Pick a graphics card (NVIDIA, AMD, Intel), an Apple chip with its real memory sizes, or CPU only with your RAM and RAM type | Included |
| See 13 open models, from 1B to 120B parameters including mixture-of-experts, each with the best quantisation that fits and an estimated speed | Included |
| Fast, Usable, Slow or Won't fit for every model, with where the model lives: all in GPU memory, split with system RAM, or RAM only | Included |
| Set the context length from 4K to 128K tokens and see the memory cost of the cache | Included |
| Reverse mode: pick a model, quantisation and context, and see the memory it needs and which graphics cards and Macs hold it entirely | Included |
| Tap a model to see the weights, cache and buffer sizes added up, so you can check the arithmetic | Included |
| Runs in your browser; nothing is uploaded | Included |
Questions
Is Local AI Fit free?
Yes. Local AI Fit is free to use with no account and no install.
Does Local AI Fit upload my data?
Nothing leaves your browser. Your choices are kept in this browser's local storage so the page reopens as you left it. The page only sends anonymous usage counts (which hardware type and model you looked at) to our analytics. See the Privacy Policy for retention and deletion.
Do I need to install anything?
No. Open the page and start working. It works on desktop and mobile browsers.
Are the speeds real benchmarks?
No. They are estimates: memory bandwidth divided by the bytes read per generated token, multiplied by an efficiency factor of 0.65. Real results vary with the software (llama.cpp, Ollama, LM Studio, MLX), drivers, batch size and cooling, but the estimate is good for comparing options and for telling usable from unusable.
What is quantisation and which one should I pick?
Quantisation stores each weight in fewer bits so the model needs less memory. Q4_K_M, at about 4.85 bits per weight, is the common balance of size and quality. Q8_0 is almost identical to the original but about 75% larger. Below Q4 the quality loss becomes noticeable; a smaller model at Q4 or higher is usually the better trade than a bigger one at Q2.
Why does context length change the memory needed?
Every token in the conversation stores a key and a value in every layer. That cache grows linearly with context length. For Llama 3.3 70B it is about 2.5 GB at 8K tokens and about 10 GB at 32K tokens, on top of the weights.
What about AMD cards, Intel Arc and Apple silicon?
They are included with their spec-sheet memory sizes and bandwidth. The estimate assumes software that can use the card well (llama.cpp with Vulkan or ROCm, or MLX on Apple). Support and speed differ between programs, so treat non-NVIDIA numbers as slightly more optimistic.
Does a model that spills into system RAM still work?
Yes, but every token has to read the part in RAM, which is several times slower than reading it from the card. A model that is a little too big for the card still runs acceptably; one that is far too big drops to a few tokens per second.