46 tps is where a local model stops feeling like a compromise. Which quant are you running to get there, and is the jump from OpenCode mostly prompt caching or better batching?