Qwen3 Coder hits 1,207 tok/s

The bigger story is not just that Qwen3 Coder runs fast on Blackwell GPUs, but that someone finally packaged the tuning process into shareable recipes instead of one-off forum screenshots.

• –vLLM crushed SGLang on the 5090 AWQ setup, delivering 2.7x higher output throughput and much better TTFT for Qwen3-Coder-30B-A3B-Instruct-AWQ
• –The 32GB RTX 5090 held roughly 115K context without throughput collapse, while the 96GB PRO 6000 handled the full 262K context window with no measurable degradation
• –On PRO 6000, SGLang won the low-concurrency comparison, but vLLM pulled ahead hard at scale and peaked at 1,207 tok/s once concurrency was pushed to 40
• –DeploDock looks like the real infrastructure play here: recipes, benchmark matrices, and GitHub Actions results turn LLM serving optimization into something teams can version, review, and replicate