Solar Pro

Upstage today announced the release of a preview version of its next-generation large language model (LLM), Solar Pro. This preview, available as an open-source model with free API access, gives developers and businesses the opportunity to test and provide feedback on the model ahead of its official release in November. As the flagship model of the Solar LLM series, Solar Pro features 22 billion parameters, more than double the size of its predecessor, Solar Mini. Despite its increased size, Solar Pro is optimized to run efficiently on a single GPU, thanks to Upstage's proprietary Depth-Up Scaling (DUS) method and advanced data recipe. This innovation enables Solar Pro to deliver state-of-the-art performance while significantly reducing model size—an essential advantage in the face of rising GPU costs and supply constraints in the AI landscape. Solar Pro's advanced capabilities have translated to impressive gains in key LLM benchmarks, with an average improvement of 51% compared to Solar Mini. It achieved an accuracy score of 52.11 on the MMLU Pro benchmark, which measures multi-disciplinary language understanding and reasoning across 14 domains. Additionally, Solar Pro excelled in the IFEval benchmark with a score of 84.37, showcasing its ability to follow complex instructions with intelligence comparable to that of humans. These results surpass those of similar-sized models from leading tech companies, including Microsoft's Phi 3 Medium, Meta’s Llama 3.1 8B, Mistral NeMo 12B, and Google's Gemma 2 27B. Solar Pro even competes with much larger models that require multiple GPUs, such as Llama 3.1 70B, which has three times the parameter count. By setting a new standard in both general and specialized tasks, Solar Pro positions itself as the most intelligent and efficient LLM available on the market today.