Upstage · 2026-07-22 · major
Solar Open 2 — Upstage's 250B/15B open-weight MoE built for agentic use
Upstage releases Solar Open 2, a 250B-parameter Mixture-of-Experts model that activates 15B per token, offers a 1M-token context, and targets agentic tool-calling. Open weights and technical report ship day one on Hugging Face.
Korea's Upstage ships a 250B/15B open-weight MoE with 1M context and a hybrid-attention stack, aimed at agentic work.
Quick facts
| Maker | Upstage (South Korea) |
|---|---|
| Size | 250B total (~15B active per token, MoE) |
| Experts | 320 routed + 1 shared; 8 experts selected per token |
| Context window | 1M tokens |
| Architecture | Hybrid Attention (interleaves linear-attention + softmax GQA) |
| License | Upstage Solar License (commercial use permitted) |
| Availability | Open weights on Hugging Face + hosted playground |
| Recommended hardware | 4x NVIDIA H200 (BF16), or 2x H200 quantized |
Benchmarks
What is it?
Solar Open 2 is Upstage's new open-weight flagship: a 250B-parameter Mixture-of-Experts language model that only activates about 15B parameters per token. Weights, a technical report, and a hosted playground all shipped together on July 22, 2026, alongside a permissive-with-attribution Upstage Solar License that allows commercial use.
How does it work?
Under the hood, Solar Open 2 interleaves three linear-attention layers with one softmax GQA layer, dropping positional encodings entirely in the linear layers (NoPE). Its routed MoE picks 8 of 320 experts per token plus a shared expert, and a Korean-tuned tokenizer cuts per-sentence tokens by 50-80% versus general models. The result: 1M-token context at inference cost closer to a mid-size dense model.
Why does it matter?
Every major frontier release this month has been either closed API (GPT-5.6 Sol, Kimi K3 for now) or Chinese-only (Qwen 3.8 Max weights still pending). Solar Open 2 gives developers a working commercially-usable open-weight model in the 250B class, and marks the first time a Korean lab ships to global open-weight leaderboards with an agent-first pitch instead of a raw-reasoning one.
Who is it for?
Enterprises building agentic workflows in Korean or English who want an open-weight fallback to closed frontier models.
Frequently asked questions
- How does Solar Open 2 score on benchmarks?
- Solar Open 2 posts 86.2 on MMLU-Pro, 92.4 on LiveCodeBench, and 70.4 on SWE-Bench Verified per Upstage's technical report. On agent-focused evals it hits 80.0 on IFBench (instruction following) and 58.2 on MCP-Atlas (tool calling). Korean-language work averages 85.4, with 86.8 on the Ko-GDPval office-work benchmark.
- What license does Solar Open 2 ship under?
- Solar Open 2 uses the Upstage Solar License, which permits commercial use but requires derivative models to include "Solar" in their name and display a "Built with Solar" attribution. The full license text is bundled in the Hugging Face repo. This is more permissive than Solar Pro 2's closed API-only release but stricter than Apache-2.0.
- What hardware do I need to run Solar Open 2?
- Solar Open 2 fits on four NVIDIA H200 GPUs in BF16, or two H200s with quantization. Because only 15B of the 250B weights activate per token, the compute cost per generation stays close to a mid-sized dense model despite the flagship-scale capacity. Upstage recommends vLLM for serving.
- Can I try Solar Open 2 without downloading it?
- Upstage runs a free hosted playground at console.upstage.ai/playground/chat where anyone can chat with Solar Open 2 in the browser. The company also lists an OpenAI-compatible endpoint that plugs into tools like Claude Code and Cursor via API. Weights are on Hugging Face for local or self-hosted deployment.
- How is Solar Open 2 different from Solar Open 100B?
- Solar Open 2 is 2.5x larger by total parameters than Solar Open 100B, but its MoE routing keeps the per-token active count near the older model. Upstage rebuilt the attention stack for Solar Open 2 as a hybrid of linear-attention and softmax layers, added Korean-tokenizer improvements worth 50-80% fewer tokens per Korean sentence, and refocused the model on agentic tool-use rather than pure reasoning.
Try it
https://console.upstage.ai/playground/chat