Apple Open Models: Model List, Licenses and Articles

September 28, 2026

About Apple

Apple releases 143 models on its apple organization on Hugging Face (a Hugging Face verified organization) as outputs of its machine learning research. These are not the Apple Intelligence models that ship in its products; most are research models released alongside papers.

A recurring theme is doing more with less compute: vision encoders that compress images into fewer tokens, reading long documents by shrinking them into images, diffusion-based code generation, and fine-tuning a model on its own outputs. Rather than competing on size, the work asks how to run a model of a given size more efficiently. The base models are often other companies’ open models such as Qwen and Mistral.

What It Releases

• LensVLM-9B (September 2026): a vision-language model built on Qwen3.5-9B that skims long documents rendered as compressed images and expands only the pages it needs back to their original form (our article).
• SimpleSD (March 2026): research checkpoints for Simple Self-Distillation, which improves code generation by fine-tuning a model on its own sampled outputs, without rewards, verifiers, teacher models or reinforcement learning. There are 4B and 30B (based on Qwen3-30B-A3B) versions.
• CADD-Base-7B (April 2026) and DiffuCoder-7B (July 2025): diffusion language models for code, which fill in the whole output step by step instead of writing left to right.
• CLaRa-7B (December 2025): research on bridging retrieval (RAG) and generation by turning retrieved documents into compressed representations, based on Mistral-7B-Instruct.
• FastVLM (August 2025): vision-language models using FastViTHD, a vision encoder that compresses high-resolution images into fewer tokens, in 0.5B, 1.5B and 7B sizes. According to the model card, the smallest variant reaches its first token 85 times faster than LLaVA-OneVision-0.5B.
• Others include the image-text embedding models MobileCLIP / MobileCLIP2, Sharp for building a 3D scene from a single photo, and the image generation research STARFlow.

Licenses

LensVLM, SimpleSD, CADD, CLaRa, DiffuCoder, FastVLM, STARFlow and Sharp, which we checked, are all under the Apple Machine Learning Research Model License. Use is limited to non-commercial scientific research and academic development; commercial products and services and product development are explicitly excluded, and the same limits apply to fine-tuned derivatives. You can try them locally, but you cannot build them into tools you use for work. Read each model’s license text before use. This summary is not legal advice.

Running Them Locally

• Most of these models are 10B or smaller, a size you can try on a consumer GPU or a Mac.
• Apple publishes MLX conversions of FastVLM itself (int4, int8, fp16), and the model card points to the official repository for running them in iOS and macOS apps.
• Many research models are meant to be used with their accompanying code. LensVLM’s page expansion and the generation procedure of diffusion language models do not work just by loading the weights into a general chat tool. Check the GitHub code linked from each model card.
• There are no official GGUF builds (as of 2026-09-28).

Sources: the apple organization card; the model cards and licenses of apple/LensVLM-9B, apple/SimpleSD-30B-instruct, apple/CADD-Base-7B, apple/CLaRa-7B-Instruct and apple/FastVLM-7B; the model list of the apple organization on Hugging Face (all as of 2026-09-28).

Our Coverage and Data

Local Model Watch has published 1 article(s) on Apple’s own models, 0 on third-party fine-tunes and quantized builds based on them, and 0 from its official blog. The model list below is collected by code from Hugging Face every day. Part of our publisher index; model families link to their family pages.

At a Glance

Recent Models on Hugging Face

The publisher lists 143 models in total on Hugging Face. Below are up to 15 of the most recent generative models (text, image, video, audio and so on), the scope of this site, per organization. Classifiers, feature extractors, research components and repositories without a model card are omitted, and the publisher’s own quantized builds and helper variants are folded into the row of the original model. Parameter counts are left blank for repositories that pack weights into integer types, where Hugging Face’s count is not the real parameter count.

From the Hugging Face API. Last changed 2026-09-28 (JST). Collected by code.

Licenses of Its Own Models (as recorded in our articles)

The license can differ from model to model even within one publisher. This is not legal advice; check each model’s license text before use.

By Model Family

Articles on Its Own Models

Last updated 2026-09-28 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log and the Hugging Face API.