Upstage Releases Solar Open 2: Open-Source Foundation Model Optimized for Agentic AI
Key Takeaways
- ▸Solar Open 2 is purpose-built for agentic AI tasks with multi-step reasoning, long-context processing, and reliable tool calling capabilities
- ▸The MoE architecture activates only 15B of 250B parameters per token, reducing computational requirements while maintaining competitive performance
- ▸Released as open-source with commercial usage rights, deployable on two H200 GPUs, addressing enterprise infrastructure constraints
Summary
Upstage has released Solar Open 2, an open-weight foundation model specifically designed and trained for agentic AI applications in real-world work environments. The model employs a Mixture of Experts (MoE) architecture that activates only 15 billion of its 250 billion total parameters per token, enabling efficient inference while maintaining competitive performance across agent, coding, knowledge, and Korean-language benchmarks.
Solar Open 2 addresses the distinct computational and reasoning demands of agentic AI, which differs fundamentally from conversational chat. The model is trained on realistic agent scenarios including tool calling, coding, document work, and office tasks, with a focus on multi-step reasoning, long-context processing, and precise instruction following. It supports a context window of up to 1 million tokens and officially supports Korean, English, and Japanese.
The model runs on two NVIDIA H200 GPUs with quantization, making it practical for enterprises to deploy on their own infrastructure. Solar Open 2 is released on Hugging Face under a commercially usable license, alongside a comprehensive technical report detailing the architecture, training methodology, and evaluation conditions.
- Trained on realistic agent scenarios from real work environments, with emphasis on verification and task completion rather than plausible-looking outputs
Editorial Opinion
Solar Open 2 represents a strategic shift in foundation model design—moving from single-turn question-answering optimized models toward systems engineered for sustained, multi-step autonomous work. By prioritizing agent-specific training data and efficient inference architecture, Upstage is challenging the assumption that larger is always better and making capable AI more deployable for enterprises. The open-source release positions Korea as a serious contributor to the global agentic AI landscape, particularly for organizations seeking alternatives to closed models.
