• Text, image, sound and speech generation now use one pinned upstream llama.cpp/GGML Vulkan
  runtime. The release includes guarded PS5 support for Ternary Bonsai 2's ternary weights and
  replaces Kokoro's old prebuilt OpenCL path with a source-built Vulkan runtime.
• Added 12 ready-to-download model presets across all four modalities, including Ternary
  Bonsai 2 27B, MiniCPM5, Granite, BoHF, Stable Audio 3, ACE-Step, Kokoro and Chatterbox.
  See Supported models for the complete list.
• Added Chatterbox voice cloning with 19 language choices, custom voice samples, and voice
  recording from DualSense or a web browser.
• PS5 and web now share conversations and runtime settings. Models prepare when a conversation
  starts, both interfaces show actual backend residency, and model switching and memory release
  work across text, image, audio and speech.
• The PS5 Workspace places New conversation after the voice controls, and the default idle
  timeout before releasing a model is now five minutes.
• All 12 curated presets completed three generation rounds each on the test title, including
  checks for model switching and valid generated outputs.