What's new
- Granite 4.2 3B Q4_K_M text chat with an 8K default, preserved embedded template, and exact revision/size/SHA-256 verification. Verified with bundled llama.cpp b10603 on Mac Metal and Windows CPU/Vulkan.
- Model Hub installed-only filtering, result counts, reset filters, clearer errors/retry, accurate disk checks for installed models, safe progress indicators and keyboard/focus improvements.
- Explicit reference-voice permission acknowledgement and transcript inputs.
- Fixed Windows runtime archive extraction and cross-platform path handling.
- Reproducible artifact verification protects cached files when a replacement fails integrity checks.
Validation
Full Go tests and vet passed on Mac and Windows. All 92 desktop tests, TypeScript checks and production builds passed on both. Real Electron search/filter/reset/details/Escape checks passed on both. Python tests, candidate/production catalog validation and published catalog checksum passed; model-watch reported healthy current artifacts.
Scope and known limitations
- Granite tool calling, FIM and long-context operation are not qualified or advertised. Larger-context and additional-platform measurements remain open in #32.
- Audio8 and Pocket TTS adapters are experimental, excluded from the public model picker. Audio8 has a Windows speech-completeness failure (#30); Pocket requires user-approved gated access and inference qualification (#31). Neither provides live desktop streaming playback yet.
- Remote catalog format remains unchanged for compatibility with older clients. Granite is built into this client.
- Download the installer/bundle matching your OS and CPU. This release does not add signing/notarization; normal platform trust warnings may still apply.
Research sources, exact artifacts, measurements and remaining qualification