10-18-2026, 04:08 PM
That matches what I was seeing. I ended up in about the same place using SGLang. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.
small models can still be good