Upstage open-sources Solar Open2 250B: 1M context, 15B active
It runs on 4×H200, scores 92.4 on LiveCodeBench and 70.4 on SWE-bench Verified, and arrives beside Motif-3 in Korea's open-weight sovereign-AI push.
Copy markdown
Solar Open2 250B, weights up on Hugging Face
Upstage's new MoE is 250B total but fires just 15B active per token across 321 experts, with a native 1M-token context and English/Korean/Japanese support. It posts MMLU-Pro 86.2, LiveCodeBench 92.4, and SWE-bench Verified 70.4 — frontier-class for a model you can pull today.
You can actually run it yourself
Minimum spec is 4×H200 (8 recommended), and an INT4/NVFP4 build already runs on two NVIDIA DGX Sparks at ~15 tokens/sec via vLLM. The 1M context uses NoPE — no rotary embeddings — so long-doc and repo-scale agents skip the usual RoPE-scaling hacks.
Read the license before you fork
It's the Upstage Solar License, not Apache: any derivative must keep a 'Solar' name prefix and show 'Built with Solar' attribution. Fine for internal tools and inference, but a real constraint if you plan to rebrand a fine-tune.
Korea's sovereign-AI wave: Motif-3 Beta too
The same national push just dropped Motif Technologies' Motif-3 Beta — 314B total, ~13B active, 384 experts, 262K context — landing #34 of 577 on Artificial Analysis' index. Catch: it's a preview checkpoint under a non-commercial license, so weights are research-only until a commercial build ships.