The fastest, lowest-cost batch transcription with Qwen3-ASR · 2026-10-28
Baseten’s official October 28 announcement reports optimized Qwen3-ASR 1.7B batch transcription at a 3,128× real-time factor on hour-long audio. An associated engineer reports 7× stock-vLLM throughput on one GPU. These are attributed vendor claims with different metrics, not independently reproduced latency or cost measurements. The linked article returned only a page shell during this review, so kernel-level explanations and a normalized comparison remain unverified.