News · 2026-07-26

Upstage Ships Solar Open 2, a 250-Billion-Parameter Model That Runs on Two GPUs

Upstage has released open weights for Solar Open 2, a 250-billion-parameter model built for agent workloads that runs on two NVIDIA H200 GPUs once quantized. The Korean company activates only 15 billion of those parameters per token, publishes a full technical report alongside the weights, and describes the release as "Korea's Sovereign Foundation Model, Built for Agentic Use."

Key facts

• 250 billion total parameters, 15 billion active per token, with a context window up to one million tokens.
• Runs on two NVIDIA H200 GPUs with quantization, which Upstage frames as the point of the design.
• Weights on Hugging Face under a commercially usable custom license, with a published technical report.
• Primary source: Upstage's release post and the Solar-Open2-250B model card.

The cost structure is the argument. Upstage is unusually direct about why it built the model this way: "An agent doesn't stop at a single model call; it cycles through planning, execution, verification, and revision, so it typically uses far more tokens than chat. For an enterprise to deploy the model on its own infrastructure, the model can't be too large or too slow." A mixture-of-experts design is the standard answer - the model holds enormous total capacity but routes each token to a small fraction of it, so you pay for storage once and compute per token as though the model were fifteen billion parameters. The trick is that storage is a fixed hardware purchase and compute is a recurring bill, and agents run up recurring bills.

The long-context handling is where the engineering gets more specific. Solar Open 2 interleaves three linear attention layers for every conventional attention layer. Linear layers hold sequence information in a fixed-size state, so their memory does not grow with input length; only the minority of conventional layers maintain a growing KV cache. Upstage explains that this "avoids having every layer's KV-cache grow in proportion to context length, and reduces the memory burden of long inputs." The remaining conventional layers use no positional encoding at all, a choice Upstage says targets extrapolation to inputs longer than those the model mostly saw in training.

The training-data story is the part that deserves more attention than the architecture. Rather than scraping more question-answer pairs, Upstage built a synthesis and verification pipeline where scenarios have unambiguous correct answers, then kept only data that passed verification. For tool calling specifically, the company writes: "Unlike conventional approaches that only judge whether the model selected the right tool, we have the model execute changes in a real environment, then read the changed state back to verify the result. The point is to separate generating a plausible answer from actually completing the task." That is the difference between grading a student on whether they wrote a plausible recipe and grading them on whether the cake came out. For coding, the model worked in a terminal, wrote and ran its own tests, and the training data includes the repair loop when tests failed.

The office-work component is a reminder that not all model training aims at the same market. Upstage explicitly built scenarios around cross-checking information scattered across documents and spreadsheets, recalculating formula-driven sheets, and handling tasks given in Korean - not merely producing fluent Korean, but completing Korean-language work. Alongside the sovereign-model framing, this is a model built for a domestic enterprise market that global labs underserve.

The license needs stating plainly, because "open weights" is doing variable work across this week's releases. Solar Open 2 is downloadable and commercially usable, but under a custom Upstage license requiring derivative model names to begin with "Solar," public "Built with Solar" attribution, and inclusion of the license itself. That is more permissive than an API and less permissive than Apache 2.0. It sits in the same middle category as MiniMax's community license and Poolside's model license - a category that now covers a substantial share of what gets called open weights.

The honest caveats are the usual ones for a launch post. Every benchmark comparison is Upstage measuring Upstage, on a suite Upstage assembled, and the company's claim to compete "with leading global models across agent, Korean, knowledge, and coding benchmarks" has not been independently reproduced. The two-H200 deployment figure depends on quantization settings that will materially affect quality, and Upstage does not specify the trade-off. What is verifiable and genuinely useful is the artifact itself: full weights, a complete technical report covering architecture, training methodology and evaluation conditions, and a license a company's lawyers can actually read.

Primary source, verified: read the paper →

Ground Truth. (2026, July 26). Upstage Ships Solar Open 2, a 250-Billion-Parameter Model That Runs on Two GPUs. Ground Truth. https://groundtruth.day/news/upstage-ships-solar-open-2-a-250b-model-that-runs-on-two-gpus.html

@misc{groundtruth:upstage-ships-solar-open-2-a-250b-model-that-runs-on-two-gpus,
  title  = {Upstage Ships Solar Open 2, a 250-Billion-Parameter Model That Runs on Two GPUs},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/upstage-ships-solar-open-2-a-250b-model-that-runs-on-two-gpus.html}
}

Topics: open-weights · korea · mixture-of-experts · agents · upstage · licensing