Arc Forge Labs Docs v2026.10.2

v2026.9.20

`distil-large-v3.5` is selectable on the faster-whisper lane (English only, about one WER point better than `turbo` at the same speed).

AI agents and tools can read these release notes as plain Markdown.

This release covers 6 pull requests.

Highlights

  • distil-large-v3.5 is selectable on the faster-whisper lane (English only, about one WER point better than turbo at the same speed).
  • Transparent brand marks, a circular badge, and Store/MSIX logos rendered from the brand SVG.
  • ONNX Runtime moved to 1.30 for the CPU and CUDA lanes.
  • parakeet-tdt-0.6b-v2 stays the English default after a head-to-head with parakeet-unified-en-0.6b; the evaluation is recorded in docs/TRANSCRIPTION_PLAN.md.
  • The published npm latest package now carries the Python source, so npx [@arcforgelabs](https://github.com/arcforgelabs)/dictate install works on Linux on the stable channel.

Changes

  • distil-large-v3.5 is selectable on the faster-whisper lane (English only, about one WER point better than turbo at the same speed).
  • Transparent brand marks, a circular badge, and Store/MSIX logos rendered from the brand SVG.
  • ONNX Runtime moved to 1.30 for the CPU and CUDA lanes. The CUDA lane now uses CUDA 13 runtime wheels and needs NVIDIA driver 580 or newer; on older drivers Parakeet falls back to CPU. onnxruntime-directml is capped at 1.24.x, the last published DirectML wheel.
  • onnx-asr 0.12 and pyannote.audio 4.0.7.
  • The desktop views follow the brand-book spacing and colour rules.
  • parakeet-tdt-0.6b-v2 stays the English default after a head-to-head with parakeet-unified-en-0.6b; the evaluation is recorded in docs/TRANSCRIPTION_PLAN.md.

Fixes

  • The published npm latest package now carries the Python source, so npx [@arcforgelabs](https://github.com/arcforgelabs)/dictate install works on Linux on the stable channel.
  • Documentation no longer describes the removed cloud, Pro, or API-key surfaces.

Contribution record

CHANGELOG/records/2026.9.20.md lists every pull request in this release once, with its related issues and thanks.