I hope they will release more of them, it's fully in our interest. If you look at download counts as "popularity", Lite models are more popular than their main models. If you look at it through the lens of likes on HF, it's the main models that are more popular.

I think their very good arch annihilates a need for API hosting of small models such as Mistral-small (7B). API of the big Deepseek v2 is basically the same cost and on average across tasks it will give higher quality results. There aren't a lot of applications that would benefit from api costs cheaper than their current offerings on the main model, though their API doesn't give you any privacy and your inputs are stored forever in some database accessible to ccp. But for local users it's a difference between running the model and not running it at all.