Find the right on-device model in seconds
82 Apple Core AI models — filter by what you have and what you need, verify license and provenance before you ship.
Loading…
82 Apple Core AI models — filter by what you have and what you need, verify license and provenance before you ship.
Core AI Catalog is agent-first: every skill below is the same command whether a human types it or an AI agent calls it. No separate API to learn — the CLI and the MCP tools read the same data.
For Claude, Cursor, and any MCP-speaking client.
pip install "coreai-catalog[mcp]"For terminals and coding agents that shell out.
pip install coreai-catalogDescribe what you're building in plain language — get ranked, license-checked recommendations back.
recommend_model(
task="private OCR on iPhone",
license="likely"
)coreai-catalog recommend \
--task "private OCR on iPhone" \
--license likelyNarrow by capability, device, license, family, or modality — combine as many as you need.
search_models(
capability="chat",
device="iphone"
)coreai-catalog search \
--capability chat \
--device iphone --license likelyBefore you ship, confirm commercial use is actually clear — not just assumed.
check_license(
model_id="qwen3-0-6b"
)coreai-catalog show qwen3-0-6bSame fields, same table, two or more models — for when "which one" is the actual question.
compare_models(
model_ids=["qwen3-0-6b", "gemma-4-12b"]
)coreai-catalog compare \
qwen3-0-6b gemma-4-12bHave audio, need an image? The transform graph chains models across modalities and returns the shortest multi-hop pipeline.
query_transforms(
from="audio",
to="image"
)coreai-catalog transforms \
--from audio --to imageFound an .aimodel artifact that's missing? This is the exact draft → validate → PR sequence — from the CLI or the model-request form.
Ran a real measurement on real hardware? That's exactly the kind of evidence this catalog is built to hold onto.
All models ranked by deployment readiness score. Filter by what you have and what you need.
| # | Model | Capability | Score |
|---|
The readiness score (0–100) answers one question: can I actually ship this model? It's not a quality or accuracy metric — it's a deployment-logistics score built from 13 source-grounded factors across four categories. Every point is earned from a field already in the catalog's data, never estimated.
Can you legally get and use the artifact right now?
| Artifact available for download | +15 |
| Commercial use: likely | +10 |
Where can it actually run on-device?
| iPhone supported | +10 |
| Mac supported | +10 |
How much extra engineering does it take to run?
| Stock runtime (no custom build) | +10 |
| No custom kernel required | +5 |
| No patch required | +5 |
| No ahead-of-time compile required | +5 |
How much do we actually know about how it performs?
| Has at least one benchmark | +10 |
| Status: confirmed (vs. reported) | +10 |
| Confidence: high / medium / low | +5 / +3 / −10 |
| Maturity: stable or active | +5 |
Full computation: coreai_catalog/catalog.py · readiness_score(). Same formula runs client-side here for every card and leaderboard row.
A source-grounded registry of Apple Core AI models — provenance, licenses, benchmarks, and recommendations.
pip install coreai-catalog
coreai-catalog recommend --task "private OCR on iPhone" --license likelyfrom coreai_catalog import Catalog
catalog = Catalog.load()
catalog.recommend(task="ocr", device="iphone")Install, then add to your AI client. 16 tools for model discovery, comparison, recommendation, and multi-modal transform pipelines.
pip install "coreai-catalog[mcp]"{
"mcpServers": {
"coreai-catalog": {
"command": "coreai-catalog-mcp"
}
}
}~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"coreai-catalog": {
"command": "coreai-catalog-mcp"
}
}
}~/.cursor/mcp.json
coreai-catalog-mcpMIT License. Not affiliated with or endorsed by Apple. commercial_use fields are triage labels, not legal advice.
Core AI Catalog is community-driven. Adding a model has two lanes — which one you use depends on a single question.
.aimodel artifact already exist on Hugging Face?You only need to index it. Use Add a Model below (CLI, issue form, or fork). The catalog never converts and never hosts weights.
Convert it with coreai-fabric (see the first card) — it produces the artifact and opens the catalog PR for you.
The model has no .aimodel artifact yet? coreai-fabric is the first-party, agent-first conversion pipeline. It converts, verifies parity, publishes to your own Hugging Face, and opens the catalog PR — all from the CLI (needs macOS + the Apple toolchain).
coreai-fabric new Qwen/Qwen3-0.6B # scaffold a recipe
coreai-fabric convert qwen3-0.6b # produce the .aimodel
coreai-fabric verify qwen3-0.6b # parity gates
coreai-fabric publish qwen3-0.6b # to your own HF
coreai-fabric register qwen3-0.6b --catalog-path ../coreai-catalog # opens the catalog PRThe .aimodel artifact already exists on Hugging Face? Index it — three ways, from most to least automated:
# CLI: drafts entries, validates, opens the PR
pip install coreai-catalog
coreai-catalog contribute model --pr
# or by hand: fork, edit catalog.yaml + artifacts.yaml, then
python scripts/validate.py && python scripts/audit.pyRan a measurement on your Apple Silicon device? Benchmarks are one signed line in benchmarks.jsonl — a dedicated lane, never mixed with model PRs.
Signed submissions (Sigstore or the relay) are gated on physics-plausibility and identity, then auto-merge by verification tier; unsigned entries route to curator review. Superseded values are retained, never overwritten.
Spotted a data error, missing license, or broken provenance link? Open an issue:
Open a new issueThe community zoo (john-rocky/coreai-model-zoo) remains an indexed reference upstream — the catalog tracks its conversions, but new first-party conversions flow through coreai-fabric.
This catalog is built on public upstream sources. Full attribution in CREDITS.md.
Core AI Catalog is MIT licensed and not affiliated with or endorsed by Apple Inc. Contributions welcome via pull request.