Upstage releases Solar Open 2, a 250B open-weight model with a 1M-token context

Upstage released Solar Open 2, an open-weight model with 250 billion total parameters, on July 22. It pairs a 1-million-token context window with a memory-saving design aimed at running long-context reasoning on a handful of GPUs.

The model, formally Solar-Open2-250B, is a mixture-of-experts system that activates 15 billion parameters per token. Its hybrid attention combines grouped-query attention with linear attention for long-context efficiency, and it stores the key-value cache in only 12 of its 48 layers to cut memory use, a big factor in how much hardware it needs.

By Upstage’s account, Solar Open 2 runs on roughly four Nvidia H200 GPUs in BF16 precision, or two with quantization. It supports English, Korean and Japanese, along with tool-calling and step-by-step reasoning.

Upstage reported benchmark scores of 86.2 on MMLU-Pro, 86.3 on GPQA-Diamond, 92.4 on LiveCodeBench v6, 70.4 on SWE-Bench Verified and 95.7 on AIME2026. The figures are the company’s own and have not been independently verified.

The weights ship under the Upstage Solar License, which requires derivative models to carry a “Solar” name prefix and a “Built with Solar” attribution, a lighter touch than a fully permissive license but with branding strings attached.