Technology Tool
Find the right fit.
Not a ranking. A way to see which local AI model fits your machine and your job, and what to expect before you download it.
1. Your machine and your task
Memory means the RAM (or GPU memory) you can give to the model.
2. What that context means
Rough sizes. Tokens are pieces of words, so the numbers depend on the language and the text.
32,768tokens in the window
24,576 wordsin English, about 0.75 words per token
49 pagesat about 500 words per page. Indonesian fits less: about 19,700 words
3. Models that fit
Memory = model weights + context memory (KV cache) + about 1 GB runtime overhead. All figures are estimates: real use varies with the app, quantization format and settings. KV cache = 2 × layers × KV heads × head size × tokens × 2 bytes. Layers, KV heads and parameters come from each model's own config on Hugging Face, checked on 4 Nov 2026. Mixture-of-experts models (A3B) still need all weights in memory; they just run faster.