Solar Open 2 Tested: 250B Open-Weight MoE Built for Agentic Work

The Architecture Trick: Three Linear Layers per Attention Layer

Most models use traditional softmax attention for everything, which gets brutally expensive as context grows. Solar Open 2 flips the ratio: for every one traditional attention layer it stacks three lightweight linear attention layers instead. Because those linear layers naturally track token order on their own, the team removed positional encoding entirely — an approach they call NoPE.

The payoff is a KV cache roughly a quarter the size of a comparable model, which is what makes a 1M-token context window practical rather than a GPU-budget disaster.

Key Takeaways

• Specs: 250B total parameters, 15B activated, mixture-of-experts, up to 1M tokens of context. Open weights, with a free hosted platform you can log into and try.
• Vendor benchmarks — grain of salt applies. The published numbers put it above much larger open models on reasoning, math, and coding, with the clearest separation on agentic tasks: tool calling, multi-step execution, and MCP. Korean-language performance is claimed as class-leading.
• AWS analysis test — the strongest result. Given a messy fictitious enterprise AWS bill (over $2M/year) and asked to act as a solutions architect, it worked service by service, checked pricing with tools, and returned genuine root-cause analysis rather than a list of big numbers. It quantified every fix in dollars, ranked them by impact versus risk, and pushed back on two obvious savings moves because they would create SLA or operational risk.
• It argued with the brief. Asked for a 30% cost cut, it came back saying the plan actually hits 40–45% — and showed the math. Mirza's read: this is the tell that separates a grounded model from a generic one, and it produced in one prompt what a paid advisory engagement might take weeks to deliver.
• Coding test — clean pass. A single-file, zero-dependency interactive "how long to reach the Moon" comparison chart with live interactivity, animations, responsive layout, and accurate physics. Output ran first try: dark space theme, star field, animated vehicles sliding across the track, editable distance and speed inputs, and time formatting that switches between hours, days, and years. One "show calculations" control didn't respond when clicked.
• Multilingual test — the weak spot. Asked for original, culturally appropriate short quotes across many languages, the results were mixed. Vocabulary and language understanding were clearly present, but some output looked like patched-together words, including at least one term the reviewer didn't recognize. His verdict: multilinguality is probably not its strongest area outside Korean.
• Thinks in Korean. Chain-of-thought was visibly running in Korean during the tests. If the interface loads in Korean, switch to English in the top-right settings.
• Interface caveat: the hosted console was described as flaky and patchy during the session.

The License Catch

The license is essentially Apache 2.0 with one extra condition. Commercial use is fully allowed, but if you fine-tune or build a derivative model on top of Solar Open 2, your model's name must start with "Solar" and you must publicly display "built with Solar". You can't strip the Upstage branding from anything you train on top of it.

If you're evaluating open-weight models for an agent stack, that naming clause is worth reading before you commit — it propagates to anything downstream of a fine-tune. Compare options on our benchmarks page.