10-18-2026, 09:40 PM
That matches what I was seeing. The pattern around speculative decoding has been pretty similar for me. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
routing beats arguing over one perfect model