Wishlist

[iplayfast]

I assume that Nexa-sdk issues are also Nexa issues? My issues are with nexa and there is no separate github for it.

1. When inferring a model that is not present and must be downloaded. Check to see if the disk space is available for it, before starting to download it, perhaps question the user if there is a problem.
2. Allow the user to determine where to store the models. Perhaps a ~/.local/share/nexa setting?
3. Allow importing of models from ollama, if the user already has them downloaded, perhaps just a link to the file?
4. Make nexa opensource. Over decades of experience with the latest and greatest, closed source tends to die when the money goes away. Opensource manages to carry on. It's easier to adapt something that I feel will be around as long as I can compile it.
5. Think models are a curiosity, but not actually useful when you are just trying to get information. I would like /nothink /noshowthink /think flags added so the amount of text displayed can be controlled. (another setting for the ~/.local/share/nexa)
6. When downloading a model it asks the type of model that it is. The list of options, uses terms I'm not familiar with. Perhaps have an English thing explaining what it is, and what to look for in the model card.

I've been having fun with it, but the opensource thing is a killjoy.

[RemiliaForever]

1. Good point, we will add it, #866.
2. This is now tracked in issue #863.
3. It is included in our roadmap, #867.
4. cc @zhiyuan8
5. Adding --enable-think=false allows disabling think, and --hide-think can be used to hide think. Run nexa infer --help for more details.
6. Most models from NexaAI can automatically detect the model type, but community models do not support this feature. Our documentation includes some basic explanations and will be further improved in the future.