UniLM (2023 predictive text model)

See original posts about this model:
https://jackcook.com/2023/09/08/predictive-text.html
https://github.com/jackcook/predictive-spy

OpenELM (2024 On-Device model)

Source: Apple State of the Union Jun/2024

They may be four years behind GPT-3, but Apple has finally offered integrated AI—including a clunky integration of ChatGPT that asks permission before proceeding—to its users on iOS and Mac. Apple’s annual Worldwide Developers Conference (WWDC) revealed ‘Apple Intelligence’ or extended AI functionality to most users of its 2.2 billion active Apple devices.

The on-device model is 3B parameters using GQA and LoRA (Apple, 10/Jun/2024). It is most likely a model called OpenELM 3.04B trained on 1.5T tokens, documented by Apple in Apr/2024. MMLU=26.76.

Read the OpenELM paper: https://arxiv.org/abs/2404.14619

View the OpenELM repo: https://huggingface.co/apple/OpenELM-3B-Instruct

See it on the models table: https://lifearchitect.ai/models-table/

Apple revealed some very limited rankings (and only bfloat16 precision evaluations) for both models:

Both the on-device and server models use grouped-query-attention. We use shared input and output vocab embedding tables to reduce memory requirements and inference cost…

For on-device inference, we use low-bit palletization, a critical optimization technique that achieves the necessary memory, power, and performance requirements. To maintain model quality, we developed a new framework using LoRA adapters that incorporates a mixed 2-bit and 4-bit configuration strategy — averaging 3.5 bits-per-weight — to achieve the same accuracy as the uncompressed models.

…the ~3 billion parameter on-device model, the parameters for a rank 16 adapter typically require 10s of megabytes. The adapter models can be dynamically loaded, temporarily cached in memory, and swapped — giving our foundation model the ability to specialize itself on the fly for the task at hand while efficiently managing memory and guaranteeing the operating system’s responsiveness.

Source: Apple State of the Union Jun/2024

The OpenELM dataset is:

Source: OpenELM paper, p2, Table 2 Dataset used for pre-training OpenELM.

• RefinedWeb: 665B
• Dolma: 580.3B (411B + 89B + 70B + 6B + 4.3B)
• RedPajama: 332B (59B + 26B + 28B + 24B + 20B + 175B)
• The Pile: 207B
• Total: 1784.3B

Read more about the contents of these datasets in my 2022 What’s in my AI? paper.

As an interesting sidenote:

• A Google query for “OpenELM” within 24h of Apple WWDC returned just 156 results.
• A Google query for “GPT-4o” within 24h of OpenAI Spring Update event returned 83,400,000 results.

Server-based model (2024)

The server-based model is possibly a version of Apple’s Ferret (Oct/2023) and Ferret-UI (Apr/2024), both based on Vicuna 13B, a Llama-2 derivative with a ‘commercial-friendly’ license covering less than 700M users only. Any legal agreements between Apple and Meta would be behind closed doors, but it certainly makes me wonder…

Read the Ferret paper: https://arxiv.org/abs/2310.07704

Read the Ferret-UI paper: https://arxiv.org/abs/2404.05719

View the Ferret repo: https://github.com/apple/ml-ferret

The server-based model could also be Apple GPT 200B, using their Ajax framework.

[Apple] has built its own framework to create large language models — the AI-based systems at the heart of new offerings like ChatGPT and Google’s Bard — according to people with knowledge of the efforts. With that foundation, known as “Ajax,” Apple also has created a chatbot service that some engineers call “Apple GPT.”

Ajax was first created last year (2022) to unify machine learning development at Apple, according to the people familiar with the effort.

Read more via Bloomberg (19/Jul/2023): https://archive.md/f3C0r

Comparisons with other LLMs

2024 optimal LLM highlights

Older bubbles viz

Feb/2024

Nov/2023

Download source (PDF)
Permissions: Yes, you can use these visualizations anywhere, please leave the citation intact.

Mar/2023

Download source (PDF)

Apr/2022

Download source (PDF)

Models table

Dr Alan D. Thompson is an AI expert and consultant, advising Fortune 500s and governments on post-2020 large language models. His work on artificial intelligence has been featured at NYU, with Microsoft AI and Google AI teams, at the University of Oxford’s 2021 debate on AI Ethics, and in the Leta AI (GPT-3) experiments viewed more than 4.5 million times. A contributor to the fields of human intelligence and peak performance, he has held positions as chairman for Mensa International, consultant to GE and Warner Bros, and memberships with the IEEE and IET. Technical highlights.