@TokenSpeed

TokenSpeed, an open-source inference engine, achieved a record-breaking 580 tokens per second running the Qwen3.5-397B-A17B model on GPUs. The performance gain for agentic workloads comes from elimina…