Second cut of sokuji-native (supersedes native-v1.0.0, which was published earlier the same day and is kept only for the record): the C ABI (sk_*) that gives the Sokuji sidecar in-process ASR (transcribe.cpp), translation (llama.cpp) and TTS (audio.cpp, five families) on one pristine upstream ggml 0.22 — no ONNX Runtime, no per-backend engine stack.
Changes since 1.0.0: the Python binding now decodes streamed translation tokens incrementally. llama.cpp token pieces can split a multibyte UTF-8 character across two callbacks (the C ABI documents this), and 1.0.0 decoded each piece on its own, so Japanese/Chinese translation output could contain U+FFFD replacement characters mid-word. No change to the C library or the models.
Five wheels, one per SKU:
sokuji_native-1.0.1-py3-none-manylinux_2_35_x86_64.whl— Linux x86_64, Vulkansokuji_native-1.0.1-py3-none-manylinux_2_35_aarch64.whl— Linux arm64, Vulkansokuji_native-1.0.1-py3-none-win_amd64.whl— Windows x64, Vulkansokuji_native-1.0.1-py3-none-macosx_11_0_arm64.whl— macOS Apple Silicon, Metalsokuji_native-1.0.1-py3-none-macosx_11_0_x86_64.whl— macOS Intel, CPU only
GPU backends: Vulkan on Linux/Windows, Metal on Apple Silicon; every family also runs on CPU. Validated on real hardware: 5/5 TTS families (plus ASR and translation) on an NVIDIA GB10 (linux-arm64), an RTX 4070 SUPER over both linux-x64 and win-x64, and an Apple M4 (mac-arm64); the bf16 rungs were validated on the GB10 (Vulkan) and the M4 (Metal). No real M1/M2/M3 Metal hardware has run this suite yet; this release relies on the architectural Apple7/Apple6 capability gates that every Mac from the M1 on satisfies, not per-generation measurement.
Linux wheel floor: manylinux_2_35 (glibc 2.35), built on ubuntu-22.04 — covers Ubuntu 22.04, Debian 12, and RHEL 9 with margin (nothing in the tree references a glibc symbol newer than 2.34).
Install: not on PyPI — install directly from this release's asset URL, e.g.:
pip install https://github.com/kizuna-ai-lab/sokuji/releases/download/native-v1.0.1/sokuji_native-1.0.1-py3-none-manylinux_2_35_x86_64.whl
sidecar/requirements.txt pins all five URLs by sys_platform/platform_machine marker, so a normal pip install -r requirements.txt resolves the right one automatically.
Links: native/README.md · design doc
This is a prerelease by design — it must never win Electron auto-update's "latest release" lookup for the Sokuji app itself.