To receive industry-leading AI updates and exclusive content, sign up for our daily and weekly newsletters. Learn more
For fans of the HBO series Game of Thrones, the word “Dracaris” has a very special meaning: Dracaris is the word used to command a dragon to breathe fire.
While there are no literal dragons in the world of generative AI, thanks to Abacus.ai, the term Dracarys has gained meaning: it is the name of a new family of open, large-scale language models (LLMs) for coding.
Abacus.ai is an AI model development platform and tools vendor that is no stranger to naming its technologies after fictional dragons. In February, the company released Smaug-72B, a dragon from the classic fantasy novel “The Hobbit.” While Smaug is a general-purpose LLM, Dracarys is designed to optimize coding tasks.
In its first release, Abacus.ai applied the so-called “Dracarys recipe” to models in the 70B parameter class, which includes optimized fine-tuning among other techniques.
“It’s a combination of training datasets and fine-tuning techniques that will improve the coding capabilities of any open source LLM,” Bindu Reddy, CEO and co-founder of Abacus.ai, told VentureBeat. “We’ve demonstrated improvements on both Qwen-2 72B and LLama-3.1 70b.”
Gen AI for coding tasks is a growing area
The overall market for gen AI in the field of application development and coding is an area of intense activity.
A pioneer in this space was GitHub Copilot, which helps developers with code completion and application development tasks. Several startups, including Tabnine and Replit, are also building features to bring the power of LLM to developers.
And then, of course, there are the LLM vendors themselves: Dracarys offers a tweaked version of Meta's Llama 3.1 generic model, while Anthropic's Claude 3.5 Sonnet, coming in 2024, is also a popular and capable LLM for coding.
“Claude 3.5 is a very good coding model, but it's a closed-source model,” Reddy said. “Our recipes are an improvement on the open-source model, and Dracarys-72B-Instruct is the best-in-class coding model.”
The numbers behind Dracarys and its AI coding capabilities
LiveBench benchmarks of the new model show a noticeable improvement with the Dracarys recipe.
LiveBench gives us a coding score of 32.67 for the meta-llama-3.1-70b-instruct turbo model. With our tuned version of Dracarys, the performance improves to 35.23. For qwen2, the results are even better: the existing qwen2-72b-instruct model has a coding score of 32.38. Using the Dracarys recipe, that score improves to 38.95.
Currently, qwen2 and Llama 3.1 are the only models that feature the Dracarys recipe, but Abacus.ai plans to add more models in the future.
“We will also be releasing Dracarys versions for Deepseek-coder and Llama-3.1 400b,” Reddy said.
How Dracarys helps enterprise coding
There are several ways developers and businesses can benefit from the coding performance improvements that Dracarys promises.
Abacus.ai currently provides Hugging Face's model weights for both its Llama- and Qwen2-based models, and Reddy said the fine-tuned models are now also available as part of Abacus.ai's enterprise services.
“This is a great option for companies that don't want to send their data to public APIs like OpenAI or Gemini,” Reddy said. “If there's enough interest, we'll also make Dracarys available in our very popular ChatLLM service for small teams and professionals.”
VB Daily
Stay up to date! Get the latest news every day by email
Thanks for subscribing! Check out other VB newsletters here.
An error has occurred.
