Baseten cuts Qwen3-TTS cost to $3 per million characters
Baseten says it has optimized open-source Qwen3-TTS to about $3 per million characters on a single H100, which by its estimate is roughly 90% cheaper than comparable closed-source text-to-speech models.
Inference provider Baseten says it has tuned the Qwen3-TTS model family to serve speech at about $3–$4 per million characters. It puts that at roughly 90% below comparable closed-source models, using a $3 per 1M estimate. The company says it runs the model for customers in voice agents, language learning and enterprise call infrastructure.
Baseten got the cost down by fitting more concurrent streams onto one replica. It priced an H100 at $6.50 an hour and benchmarked mixed-length prompts of 20–500 characters. With one stream, throughput was 110 characters per second, which works out to $16.41 per million characters, with median time-to-first-audio of 57 ms. At 25 concurrent streams, throughput rose to 594 characters per second and cost fell to $3.04, while median time-to-first-audio rose to 271 ms (431 ms at p90). At 32 streams, throughput dropped to 511 characters per second and p90 real-time factor went above 1.
The serving stack uses vLLM-Omni's disaggregated stage execution. The autoregressive "Talker," which generates acoustic tokens, and the Code2Wav audio decoder run as separate engines, each with its own scheduler. That lets the decoder work on one request while the Talker generates tokens for the next. Baseten calls this the most important lever for concurrency. It also uses CUDA graphs on the Talker, which is built on the Qwen3 LLM backbone, with graph sizes chosen from offline profiling.
Baseten says it is contributing the changes back to vLLM-Omni and will cover voice-cloning optimizations in a later post. The optimized model is available in its model library.