Qwen3.5-27B fp8, bf16 benchmark nearly ties

This reads like a strong practical case for fp8 on local coding workloads, at least on this GPU and stack. The bigger story is the low variance: the benchmark is not just saying "fp8 is close," it's saying the gap is small enough to be hard to separate from noise.

• –Ten runs per setting still did not produce a meaningful separation, which suggests precision choice may matter less than expected for this workload
• –Aider’s 224-task suite and roughly 13.3k tokens per task make this a long-context coding test, not a toy prompt benchmark
• –The setup is specific: vLLM in Podman on an RTX 6000 Pro workstation GPU, so the result may not generalize cleanly to other hardware or runtimes
• –The most interesting follow-up is longer-context testing, since fp8 KV cache often shows its seams when context gets stretched
• –For people choosing between memory savings and tiny quality deltas, fp8 looks like the default worth trying first if their use case resembles agentic coding