Add ANEForge to the model libraries registry

[sbryngelson]

This adds ANEForge to the model-libraries registry, with a matching "Use this model" snippet.

What ANEForge is

ANEForge runs computation on the Apple Neural Engine (ANE) directly, without CoreML. It compiles a lazy tensor graph into a single ANE program and dispatches it through the same daemon/kernel-driver stack Apple's own frameworks use.

It integrates with the Hub through a sentence_transformers-compatible drop-in that loads any Hub model by repo id and runs the encoder on the engine:

from aneforge.sentence_transformers import SentenceTransformer

model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")   # any Hub repo id
emb = model.encode(["Hello from the Neural Engine"], normalize_embeddings=True)

Checklist

• Open-source on GitHub: https://github.com/sbryngelson/ANEForge (MIT)
• Pip-installable: https://pypi.org/project/aneforge/ (pip install aneforge)
• Loads models from the Hub (huggingface_hub.hf_hub_download under the hood)
• Docs: https://aneforge.readthedocs.io
• Paper: https://arxiv.org/abs/2606.17090
• Org + example card on the Hub: https://huggingface.co/aneforge/sentence-embeddings

Changes

• model-libraries.ts: aneforge entry (alphabetical, between allennlp and anemoi), filter: true.
• model-libraries-snippets.ts: an aneforge snippet that loads the model with the sentence-transformers drop-in and runs .encode() on the ANE.

I matched the existing tab/quote/trailing-comma style by hand but was not able to run the full monorepo toolchain locally; happy to adjust formatting or the snippet to your conventions.

---

Note

Low Risk
Adds only static registry metadata and a display snippet; no changes to auth, downloads, or inference paths in this repo.

Overview
Registers ANEForge in the Hub model-libraries registry so models tagged with library_name: aneforge get the correct label, docs/repo links, and Use this model code on model pages.

The new entry sits alphabetically between AllenNLP and AnemoI, with filter: true so ANEForge appears in the models library filter. The snippet shows loading a Hub repo via aneforge.sentence_transformers.SentenceTransformer and running .encode() on the Apple Neural Engine (no CoreML).

^{Reviewed by Cursor Bugbot for commit b1a2df8. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit b1a2df8. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b1a2df87a9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[sbryngelson]

PR bug bot suggestions were mostly correct and now fixed. Maintainers please have a peek when you get a moment. Thanks!

[pcuenca]

Hi @sbryngelson, super cool project, congrats! 🙌

However, the Hub integration you propose is unusual. Repos are intended to host model weights that users can download and run with the associated library; in your case, the four repos you created only contain usage instructions.

I understand that ANEForge can run many existing architectures by dynamically creating a graph for the ANE, which is quite impressive. To bootstrap usage of your library, perhaps you could duplicate a few representative model repos (whose license allows) under your namespace, assign them the ANEForge library name and this way the snippets can be shown and you can get download counts. This could also grow as a community effort.

I would also recommend you include the four snippets in your GitHub's readme page, as they demonstrate a super low-friction way to test your library with many models from the Hub.

Beyond the context of this PR, I'd also be curious to know what kind of adoption and interest you are seeing. The ANE remains the unreachable but highly-efficient compute unit in Apple Silicon, and there appear to be more experimental efforts trying to leverage it through private APIs (oMLX, for instance).

[sbryngelson]

Thanks, this was helpful. I've made the changes.

The repos were unusual for the reason you spotted: ANEForge doesn't ship its own weights. It compiles an existing Hub model's architecture into an ANE program on the fly and pulls that model's weights from its original repo via huggingface_hub, so the four aneforge/* repos were instructions rather than weights. I've duplicated four permissively-licensed models under the namespace, weights unchanged, each tagged library_name: aneforge with attribution and the source license:

• https://huggingface.co/aneforge/all-MiniLM-L6-v2 (Apache-2.0, embeddings)
• https://huggingface.co/aneforge/gpt2 (MIT, text-generation)
• https://huggingface.co/aneforge/vit-base-patch16-224 (Apache-2.0, image-classification)
• https://huggingface.co/aneforge/whisper-base.en (Apache-2.0, speech-to-text)

Each loads and runs on the Neural Engine straight from its aneforge/ repo id, verified on an M5 Pro. Happy to grow this as a community effort.

Since those span four tasks, I updated the snippet in this PR to branch on pipeline_tag, so a text-generation, image-classification, or ASR model shows the right ANEForge call instead of the embeddings one.

I also added a section to the GitHub README with the four snippets, the low-friction path you pointed out: https://github.com/sbryngelson/ANEForge#run-a-hub-model-on-the-neural-engine

On adoption: still early and research-stage. The recent work has been about lowering friction, a sentence-transformers drop-in and a langchain-aneforge package on PyPI, this Hub integration, and a reverse-engineering guide of the engine on arXiv, plus a community effort to map the ANE roofline across chips. Interest so far is mostly from the Apple-silicon and reverse-engineering side rather than production users. I'm aware of oMLX and a few other private-API efforts; the shared motivation is that the ANE is the efficient unit on Apple silicon but reachable only through CoreML. I'd welcome your thoughts on where a library like this is most useful to the Hub community.

Thanks again.

[pcuenca]

Thanks for iterating @sbryngelson 🙌 This looks good to me, but let's maybe wait for another reviewer before merging.

[sbryngelson]

@pcuenca rebased on master. no other reviewers yet 😮‍💨

[pcuenca]

Thank you, Spencer!