# qwen3-tts-native

> Native Rust + CUDA streaming text-to-speech for
> Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign, designed and qualified for the NVIDIA
> DGX Spark (GB10, sm_121, linux/arm64). The whole inference path is native —
> prompt preparation, the 1.7B VoiceDesign talker, its 15-step code predictor,
> device-to-device token handoff, the incremental neural speech decoder, the
> scheduler and the HTTP server. Python, Node.js, PyTorch, SGLang and vLLM are
> not in the runtime or the image. First audio at p95 in 94 ms against 2.69 s
> for stock SGLang, at 5.68 GB peak GPU memory against 108.90 GB. Stock SGLang
> still wins aggregate throughput in every profile; that trade is the point.

Made by Luka Löhr — https://lukaloehr.com

Text plus a sentence describing a voice becomes progressive 24 kHz mono PCM.
Run it on a Spark with Docker and the NVIDIA runtime, pinned by digest:

    export QWEN3_TTS_IMAGE=ghcr.io/luka-loehr/qwen3-tts-native@sha256:...
    docker pull "$QWEN3_TTS_IMAGE"
    docker run --rm --gpus device=0 --read-only --cap-drop=ALL \
      -p 127.0.0.1:8080:8080 "$QWEN3_TTS_IMAGE"

The digest comes from the current GitHub release. A tag, a branch or `latest`
is not a deployment identity. The image carries the pinned weights, so there is
nothing to download afterwards. The runtime source is Apache-2.0.

VoiceDesign only: no voice cloning, no reference audio, no speaker enrollment.

## Setup

- [Agent setup runbook](https://qwen.lukaloehr.com/setup.txt): end-to-end setup written for a coding agent. Asks first whether they actually have a DGX Spark, makes them resolve the digest rather than a tag, then runs and verifies with real audio. Start here.
- [HTTP API](https://qwen.lukaloehr.com/api.txt): endpoints, request fields, multipart streaming semantics, audio format, languages, data handling.

## Reference

- [Performance](https://qwen.lukaloehr.com/performance.txt): the controlled native-versus-stock-SGLang study, endurance and multilingual runs, and what the numbers are not.
- [Architecture](https://qwen.lukaloehr.com/architecture.txt): the native path, warm-up and readiness gate, target hardware, and what is and is not in the image.
- [Everything in one file](https://qwen.lukaloehr.com/llms-full.txt): all of the above concatenated.

## Source

- [GitHub repository](https://github.com/luka-loehr/qwen3-tts-native): full source, docs, benchmark evidence, SBOM. Apache-2.0.
- [Releases](https://github.com/luka-loehr/qwen3-tts-native/releases): the immutable GHCR digest to deploy.
- [Research paper](https://github.com/luka-loehr/qwen3-tts-native/blob/main/research/paper/qwen3-tts-native-paper.pdf): methodology, energy results, limitations.
- [Upstream checkpoint](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign): Qwen3-TTS-12Hz-1.7B-VoiceDesign.