35 Billion Parameters with MoE Architecture for Enhanced Efficiency
Runs on a Single GPU When Quantized

Upstage announced on October 1st that it has unveiled its latest large language model (LLM), "Sola Mini 4."


Sola Mini 4 is part of the compact model lineup of Upstage’s Sola LLM and achieves both high performance and efficiency despite its small scale. By applying a Mixture of Experts (MoE) architecture to a compact model with a total of 35 billion parameters, only 3 billion parameters are selectively activated during actual inference, significantly improving efficiency.


Upstage announced on the 1st that it has released the latest large language model (LLM), 'Sola Mini 4'. Upstage

Upstage announced on the 1st that it has released the latest large language model (LLM), 'Sola Mini 4'. Upstage

View original image

In terms of performance, Sola Mini 4 scored an overall 24 points based on the Artificial Analysis Intelligence Index (AAII) benchmark, a global AI evaluation agency. This is a strong result among compact models with 3 billion active parameters, outperforming models in the same class such as Alibaba’s "Qwen 3.6-35B-A3B" and NVIDIA’s "Nemotron 3.5 Lightning." It also demonstrated comparable or even superior performance to higher-tier models with larger parameter counts, including NVIDIA’s "Nemotron 3 Ultra," Google’s "Gemma 4 31B," and Xiaomi’s "MiMo V2.5."


Additionally, Sola Mini 4 can be operated using only a single graphics processing unit (GPU) when quantization is applied, making it suitable for on-premise environments where companies install and manage the model directly on their internal servers. Because it requires fewer active parameters, the model offers faster response times and lower processing costs, making it an ideal on-premise solution for companies and institutions with high security requirements.


With this release, Upstage now offers Sola LLM in two product lines: "Pro" and "Mini." Sola Pro is specialized for complex agent tasks that demand advanced judgment, while Sola Mini is optimized for quickly and efficiently handling repetitive tasks such as step-by-step process automation. Companies can combine the two product lines based on task complexity and workload to optimize both performance and cost.


Sola Mini 4 will also be publicly released as an Application Programming Interface (API) through the global API routing platform OpenRouter. Sola Pro 4, which became the first Korean model to be officially registered on OpenRouter in August, previously ranked third for weekly token usage among closed models on the platform.



Seonghun Kim, CEO of Upstage, commented, "Sola Mini 4 leads all competing models in its class, establishing a new standard for compact LLMs. By systemizing Sola LLM with Pro for complex agent work and Mini for large-scale automation, we aim to help companies select the AI most suited to their needs."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing