Allen Institute for AI unveils OLMo model for AI development

In a significant stride toward transparency in artificial intelligence, the Allen Institute for AI (AI2) has launched OLMo, a large language model (LLM) challenging the closed-source norm. Founded by Microsoft’s Paul Allen in 2014, AI2 aims to provide a genuine alternative to restrictive models.

Unlike its predecessors,

OLMo stands out by offering a comprehensive suite of AI development tools to the public. It includes training code, training data, model weights, and evaluation toolkits, accompanied by an Apache 2.0 License. This move distinguishes OLMo from models like OpenAI’s GPT-4 and Anthropic’s Claude.

This release coincides with a broader trend in open source AI, with other players like Mistral and Meta making strides in performance. OLMo’s unique selling point lies in its ‘completely open’ tools, providing transparency in pretraining data, training code, model weights, and evaluation.

Hanna Hajishirzi, OLMo project lead, emphasised the significance of transparency, stating, “Without access to training data, researchers cannot scientifically understand how a model is working.”

Nathan Lambert, an ML scientist at AI2, pointed out OLMo’s potential, stating, “OLMo will represent a new type of LLM enabling new approaches to ML research and deployment.”

The response from the open source AI community has been positive. Jonathan Frankle, chief scientist at MosaicML and Databricks, called OLMo’s release “a giant leap for open science.” Meta chief scientist Yann LeCun echoed this sentiment, noting, “Open foundation models have been critical in driving innovation around generative AI.”