{"full_name":"apple/coreai-models","owner":"apple","name":"coreai-models","shard":"1","url":"https://github.com/apple/coreai-models","description":"Model export recipes, Python primitives, and Swift runtime utilities for on-device AI","homepage":"","topics":[],"tag_pages":[],"language":"Swift","license":"BSD-3-Clause","avatar":"https://avatars.githubusercontent.com/u/10639145?v=4","created_at":"2026-06-08T04:32:27Z","pushed_at":"2026-11-14T05:45:15Z","metrics":{"stars":2347,"forks":264,"watchers":25,"issues":188,"prs":341,"releases":2,"issues_open":92,"issues_closed":96,"prs_open":14,"prs_closed":43,"prs_merged":284,"discussions":0},"growth":{"day":{"stars":0,"forks":0,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-15","2026-11-16"],"stars":[2347,2347],"forks":[264,264],"watchers":[25,25],"issues_open":[92,92],"prs_open":[14,14]},"boards":{},"same_owner":[{"full_name":"apple/container","name":"container","shard":"1","boards":{}},{"full_name":"apple/containerization","name":"containerization","shard":"1","boards":{}},{"full_name":"apple/foundationdb","name":"foundationdb","shard":"1","boards":{}},{"full_name":"apple/password-manager-resources","name":"password-manager-resources","shard":"1","boards":{}},{"full_name":"apple/swift-certificates","name":"swift-certificates","shard":"1","boards":{}},{"full_name":"apple/swift-protobuf","name":"swift-protobuf","shard":"1","boards":{}}],"issues":[{"number":387,"title":"ANE compile failure crashes the app with an MPSGraph assertion instead of returning an error","comments":9,"created_at":"2026-10-27T23:17:49Z"},{"number":341,"title":"Streaming detokenization is O(N²): it re-decodes the whole token sequence every step","comments":9,"created_at":"2026-10-11T04:27:25Z"},{"number":406,"title":"iOS static-shape Qwen3-0.6B export no longer specializes onto the NDX path after the coreai-core 1.0.0b3 / coreai-torch 0.4.3 bump: first token 122 ms → 2.8 s on iPhone","comments":6,"created_at":"2026-11-03T09:25:03Z"},{"number":411,"title":"Model Request: NVIDIA Nemotron-3-Diarization","comments":5,"created_at":"2026-11-04T08:00:00Z"},{"number":59,"title":"Model request: MiniCPM-V 4.6 (compact vision-language model) for macOS and iOS","comments":5,"created_at":"2026-06-20T08:21:40Z"},{"number":49,"title":"Swift package fails to compile for the iOS Simulator (no such module 'CoreAI')","comments":5,"created_at":"2026-06-16T05:27:13Z"},{"number":371,"title":"ModelScope as an alternative model download backend","comments":4,"created_at":"2026-10-23T02:47:54Z"},{"number":312,"title":"[Swift runtime] A failed GPU command buffer is swallowed: the call returns normally and the pipelined engine streams <pad> tokens","comments":3,"created_at":"2026-10-01T09:32:41Z"}],"discussions":[],"latest_release":{"name":"v1.0.0","tag":"1.0.0","published_at":"2026-10-12T23:40:49Z","body":"# Core AI Models 1.0.0\r\n\r\nFirst stable release of Core AI Models: export recipes, Python primitives, and Swift runtime utilities for building on-device AI with [Core AI](https://developer.apple.com/documentation/coreai).\r\n\r\n**Requirements:** macOS and iOS 27.0+, Xcode 27.0+.\r\n\r\n## Highlights\r\n\r\n- **Model export:** Recipes to export open-source models from Hugging Face and other sources to standalone `.aimodel` files. Run `uv run coreai.model.registry --list-models` to see what's supported.\r\n- **Compression presets for macOS and iOS:**\r\n  - **macOS:** INT4 weight-only (default), and INT4 with INT8 KV cache.\r\n  - **iOS:** 4-bit palettization with channel group size 32 or 8.\r\n  - **Custom recipes:** You can pass your own [coreai-opt](https://github.com/apple/coreai-optimization) YAML config.\r\n- **Python primitives:** Building blocks for authoring custom Core AI models in PyTorch, with helpers for KV cache updates and forward methods.\r\n- **Swift runtime package (`coreai-models`):** Libraries for language models (`CoreAILM`), image and video diffusion, image and video segmentation, speech, and object detection.\r\n- **Command line tools:**\r\n  - **Language models:** `llm-runner`, `llm-server`, `llm-benchmark`\r\n  - **Vision:** `image-segmenter`, `video-segmenter`, `object-detector`\r\n  - **Generation:** `diffusion-runner`, `videodiffusion-runner`\r\n  - **Speech:** `speech-recognizer`\r\n- **Agent skills:** A plugin for Claude Code, Codex CLI, and Gemini CLI with three skills: `working-with-coreai`, `model-authoring`, and `model-compression-exploration`.\r\n\r\n## Supported models\r\n\r\n| Category | Models |\r\n| --- | --- |\r\n| Language models | Gemma 3, GPT-OSS, Mistral, Mixtral, Muse Glimmer, Phi, Qwen2.5, Qwen3, Qwen3 MoE |\r\n| Vision-language | Qwen3-VL |\r\n| Diffusion | FLUX.2, Sana Sprint |\r\n| Vision | CLIP, Depth Anything v3, EDSR, EfficientSAM, PVT v2, SAM 3, YOLOS |\r\n| Audio | CLAP, Parakeet TDT, Wav2Vec 2.0, Whisper |\r\n| Text | RoBERTa, T5 |","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}