One well known problem with OpenRouter is routing to poor quality model providers who quant the models.

So you think for example you're using Kimi k2.6....but behind the scenes, it's the 4b or 8b quantized versions.

So for open source models, I've started using the providers own service. In the case of Kimi, I don't trust any provider other than moonshot not to quant the model. So far this seems to be getting better results.

If I see a provider not specifiying if they quant or not, I assume that they do.

I'll still use OpenRouter to try new models out, but not for any real work.