Upstage Unveils Solar Mini 4, Boosting Efficiency With 3 Billion Active Parameters
TECHWORLD ·
✦ AI Summary
Upstage unveiled the small LLM Solar Mini 4 on the 1st.
Solar Mini 4 has 35 billion total parameters and a Mixture of Experts (MoE) architecture, and is designed to selectively activate only 3 billion parameters during actual inference.
Upstage said it will split its Solar LLM lineup into Pro and Mini and provide Solar Mini 4 as an API through OpenRouter.
Upstage said on the 1st that it unveiled the small LLM Solar Mini 4 and plans to strengthen its product strategy of running high-performance models alongside lightweight models.
Solar Mini 4 has 35 billion total parameters and uses a Mixture of Experts (MoE) architecture. Upstage said it was designed to selectively activate only 3 billion parameters during actual inference in order to reduce compute load and processing costs.
Upstage said Solar Mini 4 scored an overall 24 points in Artificial Analysis' Intelligence Index (AAII) v4.3.2, making it the highest score among models with around 3 billion active parameters. It cited Alibaba's Qwen 3.6-35B-A3B and NVIDIA's Nemotron 3.5 Lightning as comparison models.
Upstage also said Solar Mini 4 showed similar or better performance than NVIDIA's Nemotron 3 Ultra, Google's Gemma 4 31B, and Xiaomi's Mimo V2.5 in some evaluations. Upstage added that some of the comparison models had a larger active-parameter size than Solar Mini 4.
Upstage designed Solar Mini 4 to enhance operational efficiency. Solar Mini 4 reduces the scale of active parameters and, when quantization is applied, is designed to run on a single GPU. It can also be deployed in on-premises environments installed directly on a company's internal servers. Upstage said this is intended to improve response speed and cost efficiency for repetitive, large-volume tasks.
With the model's release, Upstage has split its Solar LLM lineup into two tracks. The lineup consists of Pro and Mini. Solar Pro is intended for agent tasks that require complex, consecutive judgments, while Solar Mini is intended for step-by-step automation and repetitive processing tasks.
Companies can combine the two models based on task difficulty and workload. The Pro model can be used for tasks that require advanced reasoning, while the Mini model can be deployed for jobs that need high-speed processing of large volumes of requests. The purpose of the Pro-Mini combination is to balance performance and cost.
Solar Mini 4 is available through the global API routing platform OpenRouter via API. In August, Solar Pro 4 was listed on OpenRouter. Upstage plans to expand its supply coverage by adding the Mini model and to broaden its touchpoints with global developers.
Upstage CEO Kim Seong-hoon said Solar Mini 4 is aimed at delivering high performance even with a small active-parameter size. He added that work types are divided into complex agent tasks and large-scale task automation, with Pro handling complex agent tasks and Mini handling large-scale task automation, and said the company intends to support AI selection based on each enterprise's task characteristics.
Source: TECHWORLD · Kim Seong-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407673
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.