Apple’s OpenELM: Transforming AI Development with a Groundbreaking Language Model

  • Apple introduces OpenELM, a Transformer-based language model.
  • OpenELM features scaled-attention mechanism for efficient parameter allocation.
  • Released framework includes data prep and training code.
  • Trained solely on publicly-available data for full reproducibility.
  • Comes in four sizes: 270M, 450M, 1.1B, and 3B parameters, each with base and instruction-tuned variants.
  • Instruction-tuned variants show 1 to 2 percentage point performance improvement.
  • Innovative layer-wise attention scaling enhances model accuracy.
  • Trained on mix of datasets including The Pile and RedPajama, with about 1.8T tokens.
  • Evaluation using LM Evaluation Harness showcases superior performance.
  • Outperforms baseline models like MobiLlama and OLMo by up to 2.35 percentage points.
  • Acknowledged by Andrew Ng’s AI newsletter, The Batch, for its capabilities.

Main AI News:

Apple has unveiled OpenELM, a groundbreaking Transformer-based language model that is set to redefine the landscape of AI development. With a focus on efficiency and performance, OpenELM boasts a scaled-attention mechanism that optimizes parameter allocation, surpassing its competitors in both efficacy and resource utilization.

In addition to the model itself, Apple has generously provided the entire framework, encompassing data preparation and training codes, to the global research community. This move marks a significant departure from traditional practices, as OpenELM was trained exclusively on publicly-available data, ensuring complete reproducibility and transparency for researchers worldwide.

OpenELM comes in four distinct sizes, ranging from 270 million to 3 billion parameters, each offering a base model and an instruction-tuned variant. Notably, Apple’s research team has demonstrated that the instruction-tuned models exhibit superior performance, boasting a 1 to 2 percentage point improvement on various benchmarks.

Speaking on the release, Apple emphasized their commitment to open research endeavors, stating, “Our comprehensive release includes not only model weights and inference code but also the entire training and evaluation framework, along with pre-training configurations and conversion code for deployment on Apple devices. This initiative aims to empower the global research community and foster collaboration in AI development.”

A standout feature of OpenELM is its innovative layer-wise attention scaling, which diverges from conventional Transformer-based architectures. By allocating fewer parameters to lower layers and increasing them in higher layers, OpenELM achieves unparalleled accuracy within a given parameter budget.

Trained on a diverse array of publicly-available datasets, including The Pile and RedPajama, OpenELM encompasses approximately 1.8 trillion tokens in its pre-training mix. For instruction-tuning, Apple leveraged UltraFeedback, a dataset comprising 60,000 prompts, employing sophisticated algorithms for optimization.

Apple’s researchers rigorously evaluated OpenELM across various tasks using the LM Evaluation Harness, demonstrating its prowess in common-sense reasoning and language understanding. Comparative analysis against baseline models such as MobiLlama and OLMo revealed that OpenELM outperformed its counterparts by up to 2.35 percentage points, despite utilizing significantly less pre-training data.

The unveiling of OpenELM has garnered attention from prominent figures in the AI community, including Andrew Ng, whose newsletter, The Batch, lauded the model’s capabilities. While acknowledging its performance on certain tasks, Ng noted areas for improvement, highlighting the complexity of mastering tasks such as MMLU (Multi-modal Language Understanding). Nonetheless, OpenELM stands as a testament to Apple’s relentless pursuit of innovation in AI research and development, setting a new standard for open collaboration and advancement in the field.


Apple’s release of OpenELM marks a significant leap forward in AI development, providing researchers with a powerful, transparent, and efficient language model. With its innovative features and superior performance, OpenELM is poised to drive advancements in various industries reliant on AI technologies, cementing Apple’s position as a key player in the field of artificial intelligence.