Which AI model runs on my machine? Choose your graphics card, your Mac, or enter your memory: the tool lists the open models that fit, with the most precise quantization possible, the memory used and an estimated speed in tokens per second. Everything is calculated in your browser.
How to use it
- Select your machine from the list, or “Other machine” to enter its memory and bandwidth.
- Indicate the number of identical cards and your system RAM: it is used to offload part of the model when the card is not sufficient.
- Choose the use case (chat, vision, reasoning, long context) and the desired context.
- Read the table: models that fit comfortably come first, then those that barely fit, then those that require spilling over to RAM, which is slow.
How memory is calculated
A model’s memory is the sum of three parts: the weights (parameters multiplied by the quantization bits, divided by 8), the KV cache, which grows with the context and the number of simultaneous requests, and an execution overhead. For each model, the tool retains the most precise quantization that fits in memory. The parameters, number of layers and KV heads come from the public configurations of the models on Hugging Face, read on a date displayed under the table, not from assumed values.
To start from a specific model rather than your machine, use the tool VRAM memory for local AI; to know if local is cost-effective, local AI or API.
Limitations to know
These figures are estimates, not calibrated on public measurements: the inference software (llama.cpp, Ollama, LM Studio, MLX, vLLM), the driver and the model format change the actual result. Speed is a range, from 50 to 70% of the memory bandwidth divided by the weights read for each token. For expert models (MoE), speed depends on the active parameters, which we only know when the model name indicates them; otherwise the displayed speed is a floor. For models with hybrid attention or sliding window, the displayed cache (indicated by “≤”) is an upper bound. Only models whose public configuration is readable are listed, and licenses are displayed as declared: check them before commercial use. The tool does not measure quality in French or actual performance on your machine.
Frequently asked questions
Why is a model that “fits” slow?
Generating a token requires rereading the active weights in memory. If part of the model is offloaded to system RAM, whose bandwidth is much lower than that of the card, speed drops: the tool indicates this with “Fits by offloading to RAM (slow)”.
My hardware is not in the list, what should I do?
Choose “Other machine” and enter the memory (in GB) and memory bandwidth (in GB/s) indicated by the manufacturer.