Every model on this site carries two separate marks. One says how capable the model is. The other says how well it runs on the hardware you are looking at. Keeping them apart stops a great model that crawls on your machine from looking like a bad one, and the reverse.
The grade: how capable the model is
| Grade | Label | Score from |
|---|---|---|
| S | Excellent | 85 |
| A | Great | 70 |
| B | Good | 55 |
| C | Fair | 40 |
| D | Basic | 0 |
The grade comes from the model’s own capability score, adjusted for the quantization or precision you would run it at. It does not depend on your hardware. A lower quantization lowers the grade a little, which is the quality cost described in quantization explained.
The Runs rating: how well it runs on this machine
We estimate generation speed from memory bandwidth and model size (why bandwidth), then compare it with a reference speed of 40 tokens per second.
- Runs great: about 30 tokens per second or more (or close to the image, video or decision speed target).
- Runs well: about 10 or more.
- Slow: about 3 or more.
- Very slow: under that.
- Won’t run: it does not fit, even with offload.
Memory status caps the result. A model that is a tight fit or needs some RAM offload tops out at Runs well, and heavy offload tops out at Slow. See tight fit and offloading.
The match score: how models are ranked for you
To sort a list for your hardware we blend four things into a score out of 100:
- Quality, up to 45 points.
- Fit, up to 25 points. A comfortable fit scores the most; offload and heavy offload score progressively less.
- Speed, up to 15 points.
- Efficiency, up to 15 points, for using your memory well rather than leaving most of it empty.
Choosing a workload such as coding adds a small bonus to models tagged for it. Hover a grade badge on any list to see its breakdown. These are estimates from specifications, not benchmarks run on your machine, so treat a close call as a tie and try both.