Qwen3 speculative decoding tops 280 tok/s on 3090

This benchmark makes a blunt point: local LLM success is mostly a systems problem now. Once the GPU is saturated, the winners are the stacks that pick the right draft model, tame chat templates, and keep formulas out of the prompt.

• –The `Qwen3-8B + 1.7B` combo is the real winner because 100% acceptance turns speculative decoding into a near-free speed multiplier rather than a fiddly optimization.
• –Qwen3.5's thinking mode is a benchmark landmine; if the serving layer doesn't cleanly disable it with `enable_thinking=false`, you're measuring a different workload.
• –The math failure is the most actionable result: every model missed the `4,811 / (1 - 0.47)` quote calculation, so pricing and margin math should stay in code.
• –The `35B-A3B`'s HVAC knowledge is real but bounded; it handled domain reasoning better than the smaller models, but the `32B` still mis-sized a garage, so scale alone isn't a substitute for judgment.
• –Cross-generation draft/target pairings are useful fallback options, but the lower acceptance rates keep same-family matches as the default sweet spot.