Descriptions:
Solar Open 2, developed by Upstage in South Korea, is a 250-billion-parameter mixture-of-experts model that activates only 15 billion parameters at inference time. Supporting up to 1 million tokens of context, it is purpose-built for agentic tasks — multi-step tool calling, MCP workflows, and long-document analysis — rather than general-purpose chat.
In this hands-on walkthrough, Fahad Mirza tests Solar Open 2 on two demanding real-world scenarios. The first is an enterprise AWS cost optimization task using a complex multi-service billing report, where the model performs root cause analysis, quantifies savings by dollar impact, ranks recommendations by risk, and even pushes back on cuts that would violate SLA commitments — behavior Mirza compares to a paid AWS consulting engagement on a $2 million-plus annual bill. The second test generates a fully interactive, zero-dependency, single-file JavaScript visualization comparing travel times to the moon across vehicles ranging from space shuttles to Usain Bolt, complete with live physics calculations and responsive animations.
The video also explains Solar Open 2’s distinctive attention architecture: instead of standard softmax attention throughout, the model uses a 1:3 ratio of traditional to linear attention layers. This eliminates the need for positional encoding entirely — a technique the team calls NoPE — and shrinks the KV cache to roughly one-quarter the size of a comparable model, which is how it achieves 1-million-token context without prohibitive GPU memory requirements.
📺 Source: Fahd Mirza · Published July 25, 2026
🏷️ Format: Review
