Considering that this model has basically zero censorship I would say it is going to take a while before companies have fine-tuned this enough before putting it in production.
Btw it is unfortunate that 'censored' became the default expectation. Mistral gives the raw model because that makes it useful for all kinds of purposes, e.g. moderation. Censorship is an addon module (or a lora or sth).
Businesses have a huge incentive to censor these models - it's appreciated that these are released without it, but a business has a lot of concerns about what is said around anything they are offering as a service.
The point was just that production implies a business use and that implies the need to make sure there are guardrails in place to make sure the model sticks to the business purpose instead of teaching people how to make pipe bombs. Not that anyone thinks that prevents people from learning how to make pipe bombs - they just don't want to be the ones doing the teaching.
To the contrary, I want need things locally/cheap/fast for generic internal business automation, so it will never generate content any outsider would ever see. Basically glue code between services with classification of data built-in.
