How to read the results
- Holds the whole model means the model and the conversation fit in the memory the machine runs AI from: a graphics card's video memory, the shared memory of a Mac, Ryzen AI Max or DGX Spark, or, on other machines, ordinary system RAM, which is the slowest of the three.
- Part in system RAM only happens with a graphics card. The model still runs, but the share that spills over runs at system-memory speed. Mixture-of-experts models such as gpt-oss and Qwen3.6 35B-A3B cope with this far better than dense models such as Llama 3.3 70B.
- Speed: about 10 tokens per second is comfortable to read as it writes; 30 or more feels instant. Below 4 is too slow for chat but can be fine for background jobs.
- Longer conversations need more memory. If a machine only just fits at 8K tokens, it may not fit a long document.
The math is explained in our bandwidth study and methodology. Not sure which model you need? Start with the best computer for local AI at every budget.