Models
Parameters
Parameters are the learned weights inside a model. The count — 7B, 70B, 405B — is a rough proxy for capability and a direct driver of cost and latency.
Also written: model parameters, model size
More parameters generally means more capability and always means more compute per token, which shows up as a higher price and slower generation.
The relationship has weakened considerably. Training data quality, instruction tuning and architecture now matter enough that a well-trained smaller model routinely beats a larger older one on practical tasks.
In practice
Parameter count stopped being a useful proxy for capability once training data quality, instruction tuning and mixture-of-experts routing entered the picture. A well-tuned smaller model beats a larger poorly-tuned one on most production tasks, and the number providers publish is increasingly not published at all.
Common questions
Does more parameters mean a better model?
Not reliably any more. Training data quality, instruction tuning and mixture-of-experts routing broke the correlation, and a well-tuned smaller model beats a larger poorly-tuned one on most production tasks.
Why do providers stop publishing parameter counts?
Partly competitive, partly because the number stopped predicting anything useful. With sparse architectures, total parameters and parameters active per token are very different figures.