Skip to content

Releases: john-rocky/coreai-kit

Release list

CoreAIKit 0.2.0

CoreAIKit 0.2.0 Latest
Latest

Choose a tag to compare

View all tags
@john-rocky john-rocky released this 10 Jul 02:44
0.2.0
12b7107

Catalog revision pinning

Every catalog entry now carries a revision — the Hugging Face commit hash of the artifact that was verified on real hardware. Every (catalog:) initializer downloads exactly that revision, multi-bundle models included (VL towers, TTS glue, OCR assets, Gemma PLE tables, ASR encoder pairs). A push to a model repo can no longer change what your app receives; pins are bumped deliberately (scripts/pin-catalog.py).

Apps on 0.1.0 are unaffected (the new field is ignored). First run on 0.2.0 re-downloads models once — bundles now cache under their revision instead of main.

Also

  • Fixed: macOS build with current Xcode 27 betas (os_proc_available_memory() is iOS-only in the newer SDK).
  • Fixed: live-catalog/built-in drift — nemotron-3.5-asr-streaming-0.6b, mineru2.5-pro, glm-ocr, and TimesFM's engine hint now ship in both; the sync test pinpoints per-entry drift and requires pins.
  • Changed: ASR / Whisper / Parakeet / Forecaster (catalog:) inits consult the live catalog like every other surface (clear unknown-id / wrong-platform errors).
  • New: CI (build + tests + pin check on every push) and a nightly end-to-end gate (pinned catalog download → GPU generate) on macOS 27 beta. See docs/STABILITY.md for the versioning and catalog contract.

Full changelog: https://github.com/john-rocky/coreai-kit/blob/main/CHANGELOG.md

Assets 2
Loading

CoreAIKit 0.1.0

CoreAIKit 0.1.0

Choose a tag to compare

View all tags
@john-rocky john-rocky released this 07 Jul 19:01
0.1.0
6a3738d

First tagged release. Pin it with SPM:

.package(url: "https://github.com/john-rocky/coreai-kit", from: "0.1.0")
  • ChatSession(catalog:) — streaming chat over 20+ hosted LLMs (Qwen3/3.5/3.6, Gemma 3/4 incl. E2B on iPhone, LFM2.5, Granite, GLM, MiniCPM, …)
  • Vision: VLM chat (Qwen3-VL, MiniCPM-V, Holo2), CLIP, depth, detection (RF-DETR/YOLOX), super-resolution, video action recognition
  • Speech & audio: ASR (Whisper / Qwen3-ASR / Parakeet / Nemotron streaming), TTS (VoxCPM / Kokoro), music generation (Stable Audio), speaker diarization + MeetingTranscriber (who-said-what)
  • Documents: OCR readers, visual document retrieval (ColModernVBERT), text embeddings, reranker, PII extraction
  • KitLanguageModel — FoundationModels provider (tool calling + guided generation) so local models drive LanguageModelSession, Spotlight tools, Siri / App Intents
  • Every model downloads from the Hugging Face Hub on first use; the live catalog updates without a package release

Requires macOS 27 / iOS 27 beta (real device). See the model zoo for cards and conversion recipes.

Loading