Larger FAST models with SO.

Currently SO by constrained decoding is supported by large models (>80B) only with tps <50. This is insanely slow for agentic architectures based on SGR, and you keep slowing down inference. Please provide at least one fast, large, good model with SO