40+tok/s - optimized recipe for Qwen 3.5 122B Int4 on a single DGX Spark with vLLM

In three linesQwen 3.5 122B Int4 optimization on single DGX Spark with vLLM achieving 40+ tokens/s. Highest speed score on spark-arena across all context lengths and concurrency levels for Int4 recipe.

Summary generated by Claude — human-verified