Baseten has improved the processing speed of the Qwen3-ASR 1.7B open-weight model to lower the cost and time required for batch transcription. The new implementation allows for the transcription of an hour of audio in less than five seconds at a cost of $0.0013 per hour, achieving 7x the throughput of stock vLLM on a single GPU.
The company achieved these gains through native audio decoding, KV cache quantization, and a reduction in CPU-GPU synchronizations. Baseten said the 1.7B parameter model currently ranks on the Pareto frontier of the Open ASR Leaderboard as the most accurate transcription model of its size.