Thanks and a few question
First of all, thanks for the conversion efforts. I downloaded this and gave it a try. It works with a custom build of mlx_lm basically, but due the model outputting custom thinking/tool markers and no parser support in most of the frontends, the usage is very limited. Opencode, Open Webui and most other tools lack support for that format, so thinking is not displayed properly and tool calling also doesn't work most of the time. I saw that there is support in vllm for Cohere's parsers and formats, but mlx_lm doesn't have anything like that. Any way this could get implemented for MLX? Some custom layer or a middleware for conversion of the markers? What would theoretically be the right way of adding the support? I'm rather new to the ecosystem, btw, which is why I'm asking.