Photo by luis gomes on Pexels

Leaked data suggests OpenAI is gearing up to launch a significant open-source AI model, potentially shaking up the landscape dominated by Meta and Mistral AI. Deleted repositories, bearing names like `yofo-deepcurrent/gpt-oss-120b` and `yofo-wildflower/gpt-oss-20b`, and including OpenAI team members in their account details, point towards a forthcoming ‘GPT Open Source Software’ offering.

The existence of multiple models hints at a strategically developed suite of AI tools. Reportedly, the larger 120-billion-parameter model leverages a Mixture of Experts (MoE) architecture, employing 128 specialized experts to process queries. This approach positions it as a potential competitor to Mistral AI’s Mixtral and Meta’s Llama models.

Further leaks indicate the potential model’s ability to handle extensive vocabularies and utilize Sliding Window Attention for processing long-form text. Releasing a `gpt-oss` model could be a strategic move by OpenAI to recapture the attention and loyalty of developers and researchers, many of whom have migrated toward the open-source ecosystems fostered by Meta and Mistral. While unconfirmed by OpenAI, the presence of leaked code and configuration files lends significant weight to the speculation. An open-source MoE model with 120 billion parameters from OpenAI would mark a pivotal moment in the evolution of AI development.