v0.8.0

What's Changed

• Add an opt-in C ABI for in-process embedding by @christopherthompson81 in #530
• Expose built-in audio utilities as a model family by @0xShug0 in #534
• Add optional server frontend pipeline by @0xShug0 in #330
• Add Niagara ASR model family by @0xShug0 in #517
• parakeet_tdt: size the encoder graph for the path run() will take by @christopherthompson81 in #536
• webui: split TTS text on sentences in Latin-script text by @christopherthompson81 in #539
• Define the task vocabulary once, and let the ABI ask about it by @christopherthompson81 in #544
• fix(models): align default context sizes in model implementations on 32-bit targets by @weidu-tpvision in #547
• Improve Higgs TTS token-limit error guidance by @mwzkhalil in #522
• Add chunked Qwen prefill with backend KV cache reuse by @0xShug0 in #549
• model_specs: declare every file the safetensors sources require by @christopherthompson81 in #550
• Add opt-in encoder-decoder framework modules by @0xShug0 in #551
• Merge YuE2 and SheetSage2 by @0xShug0 in #553
• vulkan: fix COL2IM_1D dispatch for large tensors by @AdityaAWaghmare in #542
• streaming: publish partial text as increments, through one shared publisher by @christopherthompson81 in #552
• refactor(framework): promote greedy Qwen decoder runtime and Qwen BPE bundle helpers by @gqf2008 in #555

New Contributors

• @christopherthompson81 made their first contribution in #530
• @mwzkhalil made their first contribution in #522
• @AdityaAWaghmare made their first contribution in #542

Full Changelog: v0.7.4...v0.8.0