250B total / 15B active per token (MoE)
1M tokens
Upstage Solar License
Official open-weight flagship from Upstage positioned as a Korean sovereign foundation model optimized for agentic work, long-context efficiency, and practical office and coding use rather than pure benchmark spectacle. The official model card and technical report describe Solar Open 2 as a 250B-A15B sparse Mixture-of-Experts model with 48 layers, 320 routed experts plus 1 shared expert, and a hybrid attention stack that interleaves one softmax layer with three linear-attention layers repeatedly across the network. Compared with Solar Open 1, the new release enlarges the expert pool, replaces RoPE with a NoPE design, extends context from 128K to 1M, and uses selective weight transfer from only about 2.3% of Solar Open 1’s compatible weights to accelerate training while preserving architectural freedom. Upstage says the linear-attention layers maintain sequence information in fixed-size state rather than growing KV cache linearly, enabling about one quarter of the memory and computation of an all-softmax stack at long context. The report also emphasizes that the model was pretrained on about 12T tokens with a 4:6 real-to-synthetic data mix, more than 80% English, at least 15% each of code and math, and a final 0.9T-token length-expansion phase that injects reasoning and agent data while extending context to 1M. Public benchmark tables position Solar Open 2 as a serious open-weight contender across both English and Korean tasks, including MMLU-Pro, GPQA-Diamond, HLE, LiveCodeBench, ArtifactsBench, APEX-Agents, Ko-GDPval, KMMLU-Pro, CLIcK, and other Korean reasoning and office-agent benchmarks, with claims that it roughly matches or approaches much larger models like DeepSeek-V4-Pro on some Korean agent tasks at a fraction of the size. The release also highlights support for agentic workflows, tool calling, and multilingual English/Korean/Japanese use, with public deployment guidance through Hugging Face, vLLM, and SGLang.