HyperWrite Becomes One of the First Companies to Serve a Custom Llama 3 70B

Led by CEO & Co-Founder Matt Shumer, HyperWrite has become one of the first generative AI companies to serve a custom version of Meta's Llama 3 70B model. Llama 3 is the latest iteration in the Meta Llama family of open-source large language models.

The HyperWrite team has been working for weeks to refine and optimize the Llama 3 model, resulting in a version that writes more like a human than any other model currently available. This custom Llama 3 70B has been beating every other model, including the popular GPT-4, in early user preference tests, showcasing its exceptional writing quality and speed.

One of the standout features of HyperWrite's custom model is its ability to access real-time information via function calling, which sets it apart from other offerings in the market and the standard Llama model. This capability is particularly valuable for HyperWrite users, as the platform provides AI agents, or bots, that perform tasks such as writing, research, data collection, and even software operation.

The development of this custom model is a testament to Shumer's expertise and dedication to pushing the boundaries of what is possible with open-source models. By adapting a method called rotary position embeddings (RoPE), among other techniques, and training the model on longer passages of text, Shumer and his team have successfully extended the context window of Llama 3, allowing it to process and understand more information in a single query.

While there are potential drawbacks to longer context windows, such as an increased risk of hallucination, the benefits are significant. Models with longer context windows excel at tasks like generating text in a specific author's style, as they can more precisely capture the nuances and characteristics of the given writing. This will help power HyperWrite's new Personas feature.

HyperWrite plans to fully integrate the custom Llama 3 70B model across the platform and in additional features over the coming weeks, and anticipates that it will be one of our best-performing writing models available.

The use of open-source models like Llama 3 not only enables the HyperWrite team to work with the most cutting-edge AI technology, but also significantly reduces costs compared to using closed models like GPT-4. As Shumer stated, "We've cut our costs probably at a factor of like 100. It's insane how much more cost efficient we've gotten with these models."