- Solar Open 2 is an open-weight foundation model optimized for agentic use, designed to serve as a base model for agents in real work environments.
- It uses a Mixture-of-Experts (MoE) architecture with 250B total parameters but activates only 15B per token, improving both capability and inference efficiency.
- The model supports a 1M-token context window and runs on two NVIDIA H200 GPUs with quantization, making it practical for enterprise deployment.
- It officially supports Korean, English, and Japanese, with a tokenizer optimized for Korean efficiency (50–80% of tokens compared to global models).
- Training data focused on agent scenarios including search, tool calling (MCP), coding, and office work, with a custom data synthesis and verification pipeline.
- Solar Open 2 achieved top scores among comparable models on MMLU-Pro (86.2), LiveCodeBench (92.4), and Korean benchmarks (85.4 average).
- On Korean office work benchmark Ko-GDPval, it scored 86.8, nearly matching a model six times larger (DeepSeek-V4-Pro) while using only 15B active parameters.
- The model uses Selective Weight Transfer from Solar Open 100B for efficient training, reaching target loss with 58% of the tokens needed by random initialization.
- Weights are released on Hugging Face under a commercially usable license, with a technical report covering architecture, training methodology, and evaluation conditions.
- Solar Open 2 is built for long-horizon tasks, long-context processing, and reliable tool calling, addressing the unique demands of agentic workflows.