Dolphin (Cognitive Computations)

Dolphin began with Eric Hartford, who created the Dolphin and Samantha open models and says he previously worked at Microsoft, AWS and eBay.^1, 2

He announced Dolphin on 2 July 2023 as an open-source dataset and a series of instruct-tuned language models based on Microsoft's Orca paper.^3, 8

The dataset was released under the Apache 2.0 license and was filtered to remove refusals and alignment instances, with the aim of producing uncensored models.³

The first model, dolphin-llama-13b, was first committed to Hugging Face on 2 July 2023; it is based on LLaMA 1 and therefore limited to non-commercial use.^{3, 4, 5, 6}

Later models include Dolphin Llama 3.1 8B Instruct, a retrained version of Llama 3.1 8B, and Dolphin Mistral 24B Venice Edition, built with Venice.ai and published under Apache 2.0.^5, 9

The X1 line includes an 8B model trained on Llama 3.1 8B and a 405B model trained on Llama-3.1-Tulu-3-405B.¹⁰

The group behind the original work, Cognitive Computations, is now called Quixi AI, an independent AI lab.^1, 2

The Dolphin website describes Dolphin as a project dedicated to advancing AI model development and distributed inference.⁷

The project also runs the Dolphin Network, a decentralized AI inference network, and states that it operates as a DAO with no equity structure.^7, 10

As of October 2026, Eric Hartford is presented as the founder of Quixi AI.²