Solar Open 2

Upstage's new open-weight flagship and the successor to Solar Open 100B: a 250B-total / 15B-active Hybrid-Attention MoE. Its 48 layers interleave softmax and linear attention in a [Softmax, Linear×3] ×12 pattern (75% linear), with 321 experts (320 routed, top-8, plus one shared). Because the linear-attention layers encode token order in their recurrent state, positional encoding is dropped entirely (NoPE) and only 12 of the 48 layers keep a KV cache — delivering a 1M-token context at roughly a quarter of an all-softmax model's long-context memory. Released 22 July 2026 on Hugging Face under the Upstage Solar License (commercial use, fine-tuning, and distillation permitted).

Trained cost-efficiently by selective weight transfer from Solar Open 1 (102B) — only the ~2.3% of weights that survive the architectural change are carried over, and everything else is randomly initialized. Officially supports English, Korean, and Japanese, with a Korean-optimized tokenizer that encodes text in ~50–80% of the tokens global models use.

Reported benchmarks (self-reported): MMLU-Pro 86.2, GPQA-Diamond 86.3, LiveCodeBench v6 92.4, HMMT 93.9, IFBench 80.0, MCP-Atlas 58.2, and KMMLU-Pro 78.4 — competitive with DeepSeek-V4-Flash and MiMo-V2.5 in Upstage's comparison table, and a large jump over Solar Open 100B (LiveCodeBench +36, HMMT +25). Not currently scored on the Artificial Analysis Intelligence Index.