[AI Brief] Upstage Unveils sLM “Solar Mini 4,” Strengthening Dual-Track Lineup of Pro and Mini
IT DAILY ·
✦ AI Summary
Upstage unveiled its latest language model, “Solar Mini 4,” and formally launched a dual-track product lineup alongside “Solar Pro.”
“Solar Mini 4” has 35 billion parameters, applies MoE, activates 3 billion parameters at inference time, and recorded an overall AAII score of 24, the company said.
Exem pursued the business acquisition and personnel succession of Jason, Kolon Benit announced the holding of the “Kolon Benit AX Boost Summit 2026,” and Deeplay said it supplied the predictive maintenance solution “Listen AI” for a multiaxis transfer robot.
Upstage is expanding its LLM product lineup to target demand with different work characteristics.
On the 1st, Upstage unveiled its latest in-house language model, “Solar Mini 4.”
By adding the small model “Solar Mini 4” to its existing “Solar Pro,” the company is moving to fully formalize a product lineup that divides use cases into high-difficulty agent tasks and large-scale workflow automation.
As a result, companies now have more options to match different levels of task difficulty and throughput. After introducing “Solar Pro,” Upstage is further strengthening its dual-track system with “Solar Mini 4,” allowing customers to choose based on task difficulty and throughput.
“Solar Mini 4” is Upstage Solar LLM’s small-model lineup. The model is characterized by delivering both performance and efficiency in a small model.
Its scale is 35 billion total parameters, and the architecture applies MoE. For actual inference, 3 billion parameters are activated, and Upstage said “Solar Mini 4” is notable for combining high performance and efficiency despite its small size.
Upstage said it posted an overall score of 24 on the Intelligence Index (AAII), calculated by global AI evaluation firm Artificial Analysis. This is top-tier among small models with 3 billion active parameters, and the company said it showed performance that was on par with or superior to Nvidia’s “Nemotron 3 Ultra,” Google’s “Gemma 4 31B,” and Xiaomi’s “MiMo V2.5.” Those comparison models are higher-class models with larger active parameter counts.
Upstage said this unveiling marks the full-scale launch of its Solar LLM lineup as a two-pronged system of “Pro” and “Mini.” The company explained that “Solar Pro” is for high-difficulty agent tasks involving a chain of complex decisions, while “Solar Mini” is for repetitive work such as step-by-step workflow automation, and that “Solar Mini” is characterized by fast processing and token efficiency. It also said Solar Mini 4 will be released as an API through the global API routing platform “OpenRouter.”
Upstage CEO Kim Seong-hoon said Solar Mini 4 outperforms all competing models in its class. He said Solar Mini 4 sets a new standard for small LLMs.
Kim outlined a plan to systematize complex agent tasks with “Pro” and large-scale workflow automation with “Mini.” He added that the company aims to support enterprises in selecting the most suitable AI through Solar LLM.
Photo source: Exem. Exem announced on the 1st that it had signed an agreement to acquire the business of AI security monitoring company Jason, and said it is moving to strengthen its AI business capabilities.
The transferred assets are “JSearch” and “JMachine AIOps,” and the scope includes all intellectual property rights and business rights. Exem will take over Jason’s solutions and rights.
Exem plans to integrate Jason’s technology and personnel into its existing AIOps business. Through this, it plans to advance its failure prediction and analysis capabilities and also expand its AI business.
Exem said it expects to secure additional recurring license and maintenance (MA) revenue based on the solutions it has received from Jason. Accordingly, the company explained that it expects to achieve both revenue growth and strong cash generation.
However, Exem said it does not plan to directly sell JSearch. It also said it does not plan to directly sell JMachine AIOps.
Exem decided to retain Jason’s outstanding talent as employees to ensure the stable advancement of its in-house AI solutions. Those retained through the employment succession include top-tier AI developers.
The retained employees will join Exem as of October 1. They will be assigned to the AI development organization and the field deployment engineer (FDE0) organization, where they are expected to support the company’s AI technology advancement and AI business revenue growth.
Meanwhile, Kolon Benit announced on the 1st that it will hold the “Kolon Benit AX Boost Summit 2026.” The event, named “Kolon Benit AX Boost Summit 2026,” will be held on the 13th at Seoul Dragon City in Yongsan District, Seoul. Photo source: Kolon Benit.
This year’s theme is “Customer AX Journey.” The event is aimed at sharing execution strategies for linking AI adoption to business transformation and results. Based on business domain knowledge accumulated across diverse industries, proven technology, and an ecosystem of optimal solution partners, Kolon Benit plans to present ways to support customers’ successful AX journeys.
The morning session will begin with opening remarks by Kang Ik-oo, CEO of Kolon Benit. Jang Jae-yong, head of the division, will then give a presentation on the theme “FROM AI TO VALUE: From companies adopting AI to companies creating results with AI.” Jang is expected to introduce strategies and directions for connecting enterprise AI with real work and business outcomes.
In the afternoon, 3 specialized tracks will be operated. The tracks will consist of Data & Analytics, AX Journey, and Enterprise Service. Each track will introduce technologies and solutions needed to improve decision-making using Data and AI, execution methods for AX by industry, and innovation in digital work environments.
Deeplay said on the 1st that it supplied its predictive maintenance solution “Listen AI” for a multiaxis transfer robot used by a global advanced components manufacturer.
“Listen AI” is a solution based on acoustic analysis AI that analyzes sounds from manufacturing sites in real time and provides meaningful data. The solution diagnoses abnormal sounds from the motors and reducers of multiaxis robots used for advanced parts inspection and transfer.
It also identifies the relationship between robot load and sound, tracks component degradation trends, and provides early warning indicators. Microphones installed on the robot’s drive components determine whether sounds other than normal operating noise are present, calculate abnormality scores, and include an automatic alert system.
Deeplay explained that sound-based continuous monitoring can notify users of signs of anomalies before defects occur. It added that the system records 100% of related data, contributing to data assetization.
The company also said it supports shifting from a full-line inspection method to a selective inspection method that checks only lines where abnormal signals are detected. It explained that this can help reduce maintenance labor and costs, optimize the timing of component replacement based on data, and positively affect component lifespan extension and material cost reduction.
Deeplay plans to advance sound-based predictive maintenance cases, and beyond the current stage of generating alerts when abnormal sounds are detected, it plans to realize fault-axis diagnosis and remaining useful life (RUL) prediction in the future. Based on this, the company plans to discuss expanding predictive maintenance to more lines, as well as expansion to other factories.
Source: IT DAILY · Yang Seung-gap
Original: https://www.itdaily.kr/news/articleView.html?idxno=241971
References
This article was produced with the help of an automated content generation algorithm.
Source: IT DAILY
View originalThis article was summarized and organized by BizCrush based on the original article from IT DAILY. For exact quotations and full details, please refer to the original article.