v0.1.3 linux-amd64-cuda

CUDA linux amd64 package for Qwen3-TTS native worker.

• Asset: qwen3-tts-native-linux-amd64-cuda.tar.gz
• SHA-256: eea2446480f632b64051ef2eb747cf4d62a6cadd331a0b2e0f5500ca8e711f7e
• Engine pin: 0cf4250 (Obedience-Corp/qwen3-tts.cpp)
• Verified on NVIDIA GeForce RTX 5060 Ti 16GB (driver 610.43.03)
• Platform smoke: CUDA0 talker + vocoder (REQUIRE_CUDA=1)
• Tier: 0.6B F16 default (GitHub 2GiB asset limit; 1.7B not in this tarball)
• install.json backend_hint: cuda; includes libggml-cuda.so

Does not replace v0.1.1 (CPU) or v0.1.2 (Vulkan).

v0.1.2 linux-amd64-vulkan (RADV/ANV)

Linux amd64 Vulkan package for AMD/Intel iGPU (RADV/ANV).

• Engine pin: 0cf4250 (qwen3-tts.cpp #5)
• Asset: qwen3-tts-native-linux-amd64-vulkan.tar.gz
• SHA-256: 532b72718d71dc81bc25d10badce84b8231ebef68b2753ae3b577c591c0efbf8
• install.json backend_hint: vulkan
• CPU (linux-amd64) remains v0.1.1; CUDA remains a separate -cuda lane.

obey-voice models ensure --tts selects this tarball when a hardware Vulkan ICD is present and NVIDIA is not.

v0.1.1 — native Qwen3-TTS (linux-amd64 CPU)

qwen3-tts-native v0.1.1

First linux-amd64 CPU distribution tarball. Darwin Metal remains on v0.1.0. NVIDIA CUDA is a separate linux-amd64-cuda asset when that build is published.

Assets

File	Platform	Backend	SHA-256
qwen3-tts-native-linux-amd64.tar.gz	Linux x86_64	CPU	9655d41ec044a7be43404053746c6cd0d68fb11fb0e22b3a3723aa0265c1a1fd

Layout

Schema qwen3-tts-native.install.v1. install.json uses Go's os=linux / arch=amd64 so host apps that compare runtime.GOOS/GOARCH accept the archive.

• Tier: 0.6B F16 default
• Engine pin: ed7312b8c69fcccff6285af7748fd526be2ff8cd (docs/ENGINE_PIN.txt)
• Worker: qwen3-tts-worker/v1, 24 kHz f32le, whole-utterance PCM

Smoke

GPD Pocket 4 (Ryzen AI 9 HX 370, no NVIDIA): worker ready, 57045 samples (~2.38 s) in 7.97 s elapsed. CPU is slower than realtime here; Kokoro remains the product default for conversation. Use this package for models ensure --tts on linux/amd64.

Host install

Download, verify SHA-256, unpack, run bin/qwen3-tts-worker with models/. LD_LIBRARY_PATH to bin/.

v0.1.0 — native Qwen3-TTS (darwin-arm64)

qwen3-tts-native v0.1.0

First distribution build for host apps (not tied to any single product).

Assets

File	Platform	SHA-256
qwen3-tts-native-darwin-arm64.tar.gz	macOS Apple Silicon	67b6e7514168d146cf4d892d6594559796bcea4f130c9c837beb8eac1c416e2f

Layout

See docs/DISTRIBUTION.md. Schema qwen3-tts-native.install.v1.

Notes

• Tier: 0.6B F16 default
• Backend: Metal (darwin-arm64)
• Engine pin: see docs/ENGINE_PIN.txt in the tarball / repo
• Linux/CUDA and Windows assets will land in later tags

Host install

Download the platform tarball, verify SHA-256, unpack, run bin/qwen3-tts-worker with the models/ directory. Protocol: docs/PROTOCOL.md.