Upstage has announced the release of its latest large language model (LLM), 'Solar Mini 4,' on October 1.
Solar Mini 4 is part of Upstage's small model lineup of the Solar LLM, designed to deliver high performance and efficiency even at a smaller scale. It features a total of 35 billion parameters, utilizing a mixture of experts (MoE) structure, which selectively activates only 3 billion parameters during actual inference.
The model achieved a score of 24 on the 'Artificial Analysis Intelligence Index (AAII),' a benchmark set by the global AI evaluation organization Artificial Analysis. This score is among the highest for small models with 3 billion active parameters, surpassing competitors like Alibaba's 'Qun 3.6-35B-A3B' and NVIDIA's 'NemoTron 3.5 Lightning.'
Solar Mini 4 can operate with just one graphics processing unit (GPU) when quantization is applied, making it suitable for on-premises environments where it can be directly installed on corporate servers.
With this launch, Upstage is solidifying its Solar LLM lineup into two categories: 'Pro' and 'Mini.' The 'Solar Pro' is tailored for complex decision-making tasks, while 'Solar Mini' specializes in efficiently handling repetitive tasks, such as step-by-step automation.
Solar Mini 4 will also be available via API through the global API routing platform 'OpenRouter.' To celebrate the launch, Upstage is offering a 70% discount on API pricing until October 10.
Kim Sung-hoon, CEO of Upstage, stated, "Solar Mini 4 sets a new standard for small LLMs, outpacing all competing models. We aim to help businesses choose the most suitable AI for their needs through the structured Solar LLM, with 'Pro' for complex agent tasks and 'Mini' for large-scale automation."
Meanwhile, Upstage is preparing for the final evaluation stage of a government-led independent AI foundation model project.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.