On July 23, Korea's Upstage fully open-sourced Solar Open 2 (250B-A15B). Activating only 15B parameters per token, it takes 86.8 on the Korean industrial benchmark Ko-GDPval (open-source SOTA) and, after quantization, can be deployed on just 2 H200s — pushing 1M long context into the "single-machine runnable" engineering envelope. The architecture is called Hybrid-Attention MoE: across 48 layers, every group of 4 layers, 1 Softmax layer alternates with 3 Linear layers. The Linear layers rely on recurrent hidden states to encode position order, completely removing RoPE; the entire network retains KV cache in only 12 layers, and long-context memory drops to about a quarter of a same-shape pure-Softmax model. MoE uses 320 routing experts plus 1 shared expert, with 8+1 activated per token, paired with GQA 64/8 heads. The pretraining process is also a highlight: it's initialized by selective weight transfer from the 102B Solar Open 1 — after the architecture switches from all-Softmax to Hybrid, only 2.3% of the old weights can be reused, the rest are randomly initialized, and the remaining 250B training is completed with about 2M GPU-hours (B200) plus 12T tokens. "Reusing only the compatible sub-matrix" cross-architecture transfer raises the cold-start baseline while pressing the training cost back into the "few B200 clusters can finish" range. Scores: SWE-Bench Verified 70.4, APEX-Agents 16.6 (open-source SOTA), MCP-Atlas 58.2, LiveCodeBench v6 92.4, going head-to-head with DeepSeek-V4-Flash (284B-A13B) and Mistral Medium 3.5. On the Korean side, KMMLU-Pro 78.4 and Ko-GDPval 86.8 hit open-source #1, partially exceeding Claude Haiku 4.5. Deployment-friendly: vLLM 0.22.0 has native support, the same service exposes both Anthropic-compatible and OpenAI-compatible interfaces, so Claude Code and Hermes Agent can be plugged in with zero changes; INT4 and NVFP4 quantized versions are attached, pushing the memory floor to 4 H200s. License is Upstage Solar License; derivative models need the "Solar" prefix. Judgment: linear attention's engineering value in 1M context is verified. KV cache drops to a quarter of pure Softmax, plus selective weight transfer compresses training cost back to the million-GPU-hour range — "long context + high sparse activation + 2 H200s" becomes an achievable configuration for open-source models for the first time. Solar Open 2 isn't the strongest open-source model, but it cements the line that "the strongest open source must be deployable in an enterprise's own rack".