v0.2.0

Adopting updated Core AI python packages

• Adopt latest coreai-core, coreai-torch, coreai-opt python packages from pypi (#78)

New and updated models

• Swift runner for Whisper, with autoregressive decoding and dynamic-shape tracing on export (#54, #34)
• Dynamic-shape and batch support for object detector (#29)
• Qwen3-VL export and VLM runtime support — model definition + export + runner infrastructure (#65, #68)

LLM inference engine update

• Tool calling support (#9)
• Guided generation via XGrammar (upstream Swift package) (#23, #45)
• InferenceStream with StopReason for structured generation termination (#16)
• Top-P and Min-P sampling across all engines (#48)
• Multi-turn KV cache reuse — conversation cache persists across turns (#64)
• Implicit prefix caching (#51)
• cancel() / isBusy on the InferenceEngine protocol (#32)
• Graceful exit when pointed at a model asset instead of a bundle directory (#60)

Miscellaneous

• Diffusion bundles aligned with metadata.json v0.2 schema (#33)
• Bits-Per-Weight (BPW) added to model cards (#17)
• Swift 6.2; custom AsyncSequences replace unstructured concurrency (#22, #24)
• CI (lint/build/test) on self-hosted Apple Silicon runners; GitHub issue templates (#18, #30, #26, #15)
• Optimize expert selection (#69)