qwen3-tts-native

Native packaging and runtime for Qwen3-TTS.

Ships a release tarball host apps download and unpack: a long-lived qwen3-tts-worker, GGUF models, speaker presets, and install.json checksums. The runtime is C++/GGML (Metal on macOS, CUDA/CPU on Linux)—not a Python service and not a Go library.

Org	Obedience-Corp
Engine	pinned qwen3-tts.cpp
Releases	GitHub Releases

What you get

qwen3-tts-native-<os>-<arch>.tar.gz
├── install.json          # schema qwen3-tts-native.install.v1 + hashes
├── SHA256SUMS
├── bin/qwen3-tts-worker  # product entrypoint
├── bin/qwen3-tts-cli     # one-shot smoke / debug
├── bin/libqwen3tts*
└── models/               # GGUF + presets/

Host apps verify the archive, unpack it, and start the worker. Wire protocol: docs/PROTOCOL.md. Layout and naming: docs/DISTRIBUTION.md, docs/INTEGRATION.md.

Using a release

tar -xzf qwen3-tts-native-darwin-arm64.tar.gz -C /opt/qwen3-tts --strip-components=1

export DYLD_LIBRARY_PATH=/opt/qwen3-tts/bin   # macOS; LD_LIBRARY_PATH on Linux
/opt/qwen3-tts/bin/qwen3-tts-worker /opt/qwen3-tts/models
# JSONL control on stdin, float32 PCM on stdout — see docs/PROTOCOL.md

Optional: go run ./cmd/qwen3-tts-ensure -dir <install-root> -url <tar.gz> -sha256 <hex>
(or env QWEN3_TTS_NATIVE_URL / QWEN3_TTS_NATIVE_SHA256).

Platform matrix: docs/PLATFORMS.md. Tiers: docs/TIERS.md.

Building packages (maintainers)

HF weights ──convert──► GGUF + presets
engine pin  ──build───► libqwen3tts + CLI
worker C    ──build───► qwen3-tts-worker
          ──package──► dist/*.tar.gz

Model conversion from Hugging Face is an offline maintainer step (may use Python). The published runtime is only the worker binary + GGUF artifacts.

just                         # list recipes
ENGINE_SHA=<sha> just engine pin
just engine build            # macOS Metal; Linux CPU
# Linux + NVIDIA:
#   CUDA=1 just engine build
just engine worker
just convert models
just release package
just harness smoke
just bench platform          # skips cleanly if binaries/models missing

Publishing: docs/RELEASE.md.

Repo layout

.
├── tools/           # C: worker_main.c, extract_embedding.c
├── scripts/         # convert, package, bake presets
├── third_party/     # engine pin (local checkout)
├── models/          # local GGUF + presets (not committed)
├── dist/            # release tarballs (not committed)
├── docs/            # protocol, distribution, platforms
├── cmd/             # optional Go CLIs (ensure, smoke, bench)
├── pkg/             # helpers for those CLIs
└── .justfiles/      # maintainer recipes

Go under cmd/ and pkg/ is maintainer tooling only. Hosts can integrate in any language against the worker protocol.

Status

• Stage A: warm worker, whole-utterance PCM after synth, soft cancel, 0.6B + presets
• Stage B (planned): mid-synth PCM stream + mid-synth cancel
• 1.7B: engine + convert support (talker/code_pred split + small_to_mtp); optional package when GGUF present — see docs/TIERS.md

License / models

Engine and model licenses follow upstream Qwen / qwen3-tts.cpp. Do not commit multi‑GB GGUF blobs to git.