First ggml-only sidecar bundle — replaces the ONNX Runtime / sherpa-onnx / CTranslate2 / MLX stack of the sidecar-v0.1.x line with sokuji-native (native-v1.0.1): one pristine ggml behind a single C ABI for ASR, translation and TTS.
Five bundles, one per SKU:
sidecar-linux-x64-v0.2.0.tar.zst— Vulkan (170.9 MB)sidecar-linux-arm64-v0.2.0.tar.zst— Vulkan (135.3 MB)sidecar-win-x64-v0.2.0.tar.zst— Vulkan (55.3 MB)sidecar-mac-arm64-v0.2.0.tar.zst— Metal (48.7 MB)sidecar-mac-x64-v0.2.0.tar.zst— CPU only (50.7 MB)
Contents: an embedded Python interpreter (python-build-standalone), the sokuji_sidecar package, and sokuji-native 1.0.1 built in. Every bundle in this release actually contains the native wheel — earlier sidecar-v0.1.x bundles and this refactor's own pre-release dry runs shipped hollow (no sokuji_native installed at all, so ASR/translation/TTS silently did not work); that gap is closed here, not carried forward.
Known gaps: mac-x64 ships CPU-only by design (no GPU lane is built for Intel Macs — the wheel is the none lane); real M1/M2/M3 Metal hardware has not run the TTS validation suite (only an Apple M4); moss_tts_nano's bf16-on-Vulkan output clips at a peak of 1.228, not yet root-caused (suspected sampling variance from its stop-decision sampling, ruling R23).
Install: the Electron app downloads and installs this automatically once its own sidecarVersion (package.json) is bumped to 0.2.0 — that bump ships in a normal, separate app release, not this tag. Manual inspection: https://github.com/kizuna-ai-lab/sokuji/releases/download/sidecar-v0.2.0/manifest.json.
This is a prerelease by design — it must never win Electron auto-update's "latest release" lookup for the Sokuji app itself.