Qwen3-ASR gains grassroots edge over Whisper

This is the kind of community result that matters more than a glossy launch post: developers are finding Qwen3-ASR strong enough to displace Whisper in real workloads, not just benchmarks.

• –The big story is not just raw accuracy but deployment practicality: the linked `qwen3-asr-openai` repo wraps Qwen3-ASR behind OpenAI-style transcription endpoints and realtime streaming
• –Qwen’s official materials claim strong multilingual performance, streaming support, and competitive results against Whisper-large-v3, so the Reddit test is directionally consistent with the vendor story
• –Korean performance is the standout signal here because multilingual ASR often looks good on English benchmarks and falls apart on real non-English usage
• –The post also highlights an operational caveat: vLLM looked good initially for realtime use but degraded over long sessions, while a simpler Transformers chunking setup stayed more reliable
• –If these results keep replicating, Whisper stops being the default open-source ASR choice and becomes the legacy baseline everyone now has to beat