Abacus.ai, an AI model development platform and tools vendor, has released a new family of open large language models (LLMs) for coding called Dracarys. The name comes from the HBO series Game of Thrones, where “Dracarys” is the command to a dragon to breathe fire. While there are no actual dragons in the world of generative AI, Abacus.ai has given meaning to the term Dracarys in the context of coding.
Dracarys is designed specifically to optimize coding tasks, unlike Abacus.ai’s previous model, Smaug-72B, which is a general-purpose LLM. The company has applied its “Dracarys recipe” to the 70B parameter class of models, which involves optimized fine-tuning and other techniques to improve coding abilities. According to Bindu Reddy, CEO and co-founder of Abacus.ai, the Dracarys recipe has demonstrated improvements in both Qwen-2 72B and LLama-3.1 70B.
The market for generative AI in application development and coding is seeing significant activity. GitHub Copilot, an early pioneer in the space, helps developers with code completion and application development tasks. Other startups like Tabnine and Replit are also building features that leverage the power of LLMs for developers. Additionally, LLM vendors themselves, such as Meta and Anthropic, have released models like Llama 3.1 and Claude 3.5 Sonnet, respectively, that are well-regarded for coding.
Compared to the existing models, the LiveBench benchmarks show a marked improvement in coding performance with the Dracarys recipe. The Dracarys tuned version of the meta-llama-3.1-70b-instruct model boosts the coding score from 32.67 to 35.23. Similarly, the Dracarys recipe improves the coding score of the qwen2-72b-instruct model from 32.38 to 38.95. While currently only qwen2 and Llama 3.1 have the Dracarys recipe, Abacus.ai plans to release it for more models in the future, including Deepseek-coder and Llama-3.1 400b.
The improved coding performance of Dracarys holds potential benefits for developers and enterprises. Abacus.ai provides the model weights on Hugging Face for both Llama and Qwen2-based models, making them accessible for enterprise use. Additionally, the fine-tuned Dracarys models are available as part of Abacus.ai’s Enterprise offering, providing a viable option for enterprises that prefer not to use public APIs like OpenAI and Gemini. Abacus.ai also plans to make Dracarys available on its ChatLLM service for small teams and professionals if there is sufficient interest.
