Solar Open 2 gets a 4-bit build: 250B agentic MoE, 153GB to self-host
Nota's NVFP4 quant keeps 99.7% of quality while Upstage's tech report and the Aug 8 sovereign-AI cut put the 15B-active open model in play for builders.
Copy markdown
250B total, 15B active — frontier scores, a fraction of the compute
Solar-Open2-250B is a hybrid-attention MoE (321 experts, 15B active per token) scoring 92.4 on LiveCodeBench, 95.7 on AIME2026, and 70.4 on SWE-Bench Verified. You get frontier coding and agentic numbers while paying for only 15B parameters of compute per token.
Nota's NVFP4 quant: 500GB down to 153GB, ~0.3% quality lost
A new 4-bit (W4A4) build shrinks the model 69%, from 500.6GB to 153.3GB, with the 16-benchmark average slipping just 81.57 to 81.35 (LiveCodeBench actually rose). It leans on Blackwell FP4 tensor cores at tensor-parallel-size 4, so budget for four Blackwell-class GPUs.
Tuned for agents: tool calls, code, office work, 1M context
It's trained on tool-calling, coding, and office-work scenarios with a 1M-token window, and its Ko-GDPval (86.8) nearly ties DeepSeek-V4-Pro (86.9) at one-sixth the size. The Upstage Solar License permits commercial use, fine-tuning, and distillation, so you can adapt or shrink it freely.
Build this weekend: a private, no-API-bill agent
With commercial weights, a 1M context, and quantized builds Upstage says run on as few as two H200s, you can stand up a self-hosted agent that reads long docs and drives tools with no per-token bill and data that never leaves your box. Community GGUF and INT4 builds are already up for lighter or non-Blackwell rigs.
Why now: the last finalist before Korea's Aug 8 cut
Solar Open 2 is Upstage's entry in Korea's sovereign-AI program, and the Aug 8-11 round, judged by 200 citizen evaluators on real-world agentic tasks, trims the field from four teams to three after LG's EXAONE 2.0 and SKT's A.X K2 already shipped. Expect more open weights as the deadline nears.