Core AI model zoo

Card definition line

Every model card in this zoo’s own namespaces (mlboydaisuke, coreai-community) opens with the same one-sentence definition of Core AI, placed right after the YAML front matter. The related repos (coreai-model-zoo, coreai-kit, apple-silicon-llm-bench, awesome-core-ai, coreai-assets) open with the same sentence. Adopted 2026-09-05.

The sentence

Core AI is Apple’s on-device ML runtime in iOS 27 / macOS 27 and the successor to Core ML: PyTorch models are exported with Apple’s coreai-torch (LLMs: coreai.llm.export) into .aimodel bundles that run on the GPU or the Neural Engine, e.g. Qwen3-8B 4-bit decodes at 94 tok/s on an M4 Max GPU, MLX 90 under the same protocol (apple-silicon-llm-bench, macOS 27 beta 26A5353q, 2026-06-11).

Why one line, and why this one

• A reader who lands on any single card (a person, a search engine, a coding agent reading the raw README) gets the same definition: what Core AI is, which Apple tools export to it, one measured number with its device, OS and date, and where the number comes from.
• One sentence, nothing repeated elsewhere on the card. The card’s own content is untouched below it.
• The number is one row of the bench repo’s README: Qwen3-8B 4-bit, Apple llm-benchmark vs mlx_lm benchmark (mlx-lm 0.31.3), 512-token prompt / 1024 generated / mean of 5, macOS 27.0 beta 26A5353q, 2026-06-11, raw logs under results/raw/2026-06-11-m4max-coreai-matrix/. It was chosen because it is the row whose headline, detail table and raw logs agree, on current-generation (27β) artifacts. The Qwen3-0.6B iPhone numbers depend on the export generation and were not used.
• No superlatives. The comparator (MLX 90) is stated so the sentence reads as a measurement, not a claim.

Updating it

• Edit tools/card_first_line.txt (one line). python3 tools/card_first_line.py dry-runs and prints the diff per card; --replace rewrites a card that carries an older variant (recognised by the leading “Core AI is “); --go writes one commit per card (Card: definition line (Core AI, YYYY-MM)). Update the repo READMEs by hand.
• Ownership is checked live: a card whose oldest commit is not the owner’s is excluded and gets a pull request instead. On 2026-09-05 that was Real-ESRGAN-CoreAI and Moebius-CoreAI in coreai-community (created by a contributor).
• Mirrors in coreai-community carry the sentence too, above their “Mirror of” line.
• The LiteRT cards the owner created in litert-community carry a sibling sentence (LiteRT / litert-torch, one parity number) with the same tool, its targets and sentence kept outside this repo. When the wording or the freshness rule changes, update both sets.
• When the number goes stale (new OS, new export generation), change the sentence and re-run with --replace.

This site is open source. Improve this page.