Skip to content

v0.2.0

Latest

Choose a tag to compare

@github-actions github-actions released this 31 May 23:45

[0.2.0] - 2026-05-31

Added

  • vona-core generation contracts for text token streaming plus decoupled audio transcription and synthesis traits.
  • vona-ollama loopback text adapter for Ollama /api/generate streaming, exposed through the umbrella vona/ollama feature.
  • vona-mlx Apple Silicon MLX audio engine facade, optional mlx-models text-loader hook, and umbrella MLX feature exports.
  • vona-mlx-speech shared native Rust speech model loading utilities for safetensors discovery, metadata parsing, and MLX tensor materialization.
  • vona-mlx-whisper native MLX Whisper loader and smoke example surface for local speech recognition experiments.
  • vona-mlx-qwen3-tts native MLX Qwen3 TTS loader and smoke example surface for local speech synthesis.
  • Vona model-provisioning manifests for Distil-Whisper Large V3 and the bf16 MLX Qwen3 TTS checkpoint so downstream apps can explicitly download speech assets before loading vona-mlx.
  • mlx_ollama_voice_bench example and MLX/Ollama benchmark documentation covering the local voice pipeline validation flow.

Changed

  • Split the MLX backend surface into granular contracts and crates so Ollama text generation, MLX audio orchestration, shared speech loading, Whisper, and Qwen3 TTS remain independently optional.
  • Split vona-mlx feature gates so native-mlx enables only the MLX runtime and mlx-models-loader opts into the heavier mlx-models and tokenizer text-loader stack.
  • Narrowed workspace reqwest defaults so local Ollama-only builds do not inherit TLS support, while remote HTTPS crates opt into Rustls explicitly.
  • Removed unused direct dependency surface from vona-mlx after the bloat review without changing runtime behavior.

Fixed

  • Corrected native Qwen3 TTS vocoder transposed-convolution stride inference so 12.5 Hz speech tokens decode to the expected 24 kHz waveform duration.
  • Corrected the Ollama Phi mini default tag to phi4-mini, matching the Ollama library tag accepted by ollama pull.
  • Updated benchmark notes to mark the earlier long-run Qwen3 TTS timings as historical because they were collected before the stride fix.