S-Universe releases Solar Open 2, 250B MoE with 1M-token agent context
S-Universe published the Solar Open 2 technical report for a 250B-A15B Mixture-of-Experts model built for long-horizon agentic workloads. The model expands context to 1 million tokens with a hybrid attention stack so full agent trajectories can stay in context.
Key takeaway
Open weights at this scale with million-token hybrid attention directly target the state and history bottleneck that forces aggressive summarization in multi-step agents.
Context
Solar Open 2 is positioned as an agent-first MoE rather than a general chat upgrade: the report frames long-horizon task completion as the primary design goal, with parameter count and sparse activation meant to support extended tool-using workflows.
The 1M-token hybrid attention claim matters for builders who currently truncate or summarize session traces mid-run; keeping full trajectories in-window can change how planners, critics, and tool loops are structured without inventing new memory layers.
Numbers to know
- 250B-A15BMixture-of-Experts total and active parameter scale cited for Solar Open 2
- 1 million tokensExpanded context window size via hybrid attention for agent trajectories