Are there plans to support VLM models as well. Also will custom tokens be supported in the tinker API as well ?