When Upstage released Solar Open 2 on July 22, 2026, it did not arrive as another scaled-up general-purpose language model. It arrived as something rarer: a 250-billion-parameter open-weight model that was designed from day one to run AI agents in production enterprise environments.
The benchmarks tell a specific story. Solar Open 2 scores 16.6 on APEX-Agents, the highest score in its comparison class, ahead of DeepSeek-V4-Flash (13.4) and MiMo-V2.5 (13.2). On Ko-GDPval, a benchmark for deliverable-producing Korean officework, it scores 86.8, essentially matching DeepSeek-V4-Pro (86.9), a 1.6-trillion-parameter model more than six times its size. It runs on two NVIDIA H200 GPUs with quantization applied. That combination of task-level performance and deployment practicality is what makes this release worth tracking closely.
What Solar Open 2 Is
Solar Open 2 is Upstage’s second-generation open-weight foundation model, developed in part through South Korea’s national sovereign AI foundation model program. The model uses a Mixture-of-Experts (MoE) architecture with 250 billion total parameters, of which only 15 billion are activated per token during inference. That distinction is critical for the agentic use case, which is covered in detail below.
The model supports a 1-million-token context window, achieved through a hybrid attention stack that interleaves one softmax attention layer among every three linear-attention layers. Official support covers Korean, English, and Japanese. Weights are published on Hugging Face (upstage/Solar-Open2-250B) under the Upstage Solar License, an Apache 2.0 derivative that permits commercial use and fine-tuned derivative development. A full technical report is available at arxiv.org/html/2607.20062.
The Agent-Native Training Methodology
Most large language models are trained for conversational quality, then later evaluated on agentic benchmarks as a secondary concern. Solar Open 2 inverts this. Upstage trained twelve domain specialists across purpose-built agent scenarios including tool calling, coding, office document workflows, and multi-step reasoning tasks. These specialists were then consolidated into a single model using what the company calls Multi-teacher On-Policy Distillation (MOPD).
The training philosophy flows from a practical observation: agents do not behave like chatbots. An agent completing a business workflow may run 50 to 200 inference steps, each requiring accurate tool calls, context retention across a full task history, and reliable instruction following. A model optimized for single-turn conversational quality can degrade significantly over that kind of trajectory. Solar Open 2’s training was designed around that operational reality, not around producing impressive short-context demonstration responses.
Upstage refers to this training environment as “Officeverse,” a simulation environment covering real enterprise productivity tasks: document writing, spreadsheet analysis, presentation creation, and coding. Evaluation measured task-completion accuracy on realistic deliverables, not just surface-level correctness on multiple-choice benchmarks.
Benchmark Results
The following table summarizes Solar Open 2’s performance across key benchmarks against comparably sized open-weight models and fast-tier closed APIs.
| Benchmark | Solar Open 2 (250B) | DeepSeek-V4-Flash (284B) | MiMo-V2.5 (310B) | Mistral Med 3.5 (128B) |
|---|---|---|---|---|
| MMLU-Pro (knowledge) | 86.2 | 82.4 | 83.5 | 74.6 |
| LiveCodeBench (coding) | 92.4 | 92.3 | 90.4 | 84.9 |
| IFBench (instruction) | 80.0 | 67.1 | 73.9 | 69.0 |
| MCP-Atlas (tool calling) | 58.2 | 63.9 | 30.7 | 27.2 |
| APEX-Agents (agent overall) | 16.6 | 13.4 | 13.2 | 6.1 |
| Korean avg | 85.4 | 84.9 | n/a | n/a |
| Ko-GDPval (office work) | 86.8 | 85.0 | 84.6* | n/a |
*Ko-GDPval figure for MiMo-V2.5-Pro (1T), not the 310B variant. DeepSeek-V4-Pro (1.6T) scored 86.9 on Ko-GDPval.
The most significant result is not any single benchmark number but what the combination implies: Solar Open 2 leads on APEX-Agents and IFBench, the evaluations most directly tied to how reliably an agent completes real tasks, while also being competitive with or ahead of larger models on knowledge, coding, and Korean officework. Sources: Upstage Solar Open 2 blog, Solar Open 2 Technical Report, Korea Times, Digital Today.
Why the 15B Active-Parameter Architecture Matters for Enterprise AI
This is where the business case becomes concrete.
Agentic enterprise workflows consume tokens at an entirely different rate than chat interfaces. A customer support agent that handles a single complex support ticket might run 80 to 150 inference calls across planning, tool execution, verification, and response generation steps. Each call is a separate forward pass through the model. For a dense 1-trillion-parameter model, that means 80 to 150 full passes at full cost.
Solar Open 2 activates 15 billion parameters per token, regardless of the total 250-billion scale. That is roughly one-sixth of DeepSeek-V4-Flash’s per-token active compute, less than one-hundredth of a dense 1T model. At enterprise scale, across agents running thousands of multi-step workflows per day, that difference in active-parameter cost is not marginal. It is structural.
The infrastructure floor is equally important. The BF16 (full precision) model requires four NVIDIA H200 GPUs. With standard quantization applied, it runs on two H200s. For organizations that already run GPU clusters for inference or fine-tuning, this puts Solar Open 2 within reach of an on-premise or VPC-deployed configuration rather than requiring dedicated frontier-model infrastructure contracts. Compare that with self-hosting options in the 1T-plus parameter range, which require large A100 or H100 clusters and present significant operational complexity.
The Open-Weight and Licensing Details Enterprise Teams Need
The Upstage Solar License is an Apache 2.0 derivative with one additional requirement. Derivative AI models distributed externally must have names beginning with “Solar” and display “Built with Solar” in associated interfaces or documentation. Commercial use is explicitly permitted. Fine-tuning is permitted. Distillation of Solar Open 2 outputs into smaller specialist models is permitted, which is the license right that matters most for teams building domain-specific agents for internal deployment.
This is a materially different commercial position from models with more restrictive licenses or from closed APIs where fine-tuning rights, data terms, and output-use policies require careful legal review. For enterprise legal and procurement teams that have developed fatigue around AI vendor terms, Solar Open 2 offers a relatively clean foundation.
For enterprises interested in the open-weight frontier model landscape more broadly, the Kimi K3 release earlier this month demonstrated both the potential and the compliance risks of Chinese open-weight models. Kimi K3’s IP accusation is a reminder that open weights from certain geographies now carry non-trivial procurement risk. Solar Open 2, developed in South Korea under a government sovereign AI program, sits in a different compliance risk category.
Sovereign AI and the Non-US, Non-China Option
South Korea’s national sovereign AI program funded part of Solar Open 2’s development. The explicit goal, as stated by Upstage CEO Kim Sung-hoon, is to give Korean institutions an alternative to both US commercial AI and Chinese AI infrastructure, while building toward what the company calls a “full-stack sovereign AI” combining domestic NPU chips with open models and agent services.
For enterprise buyers in regulated industries, governments, and organizations subject to data-residency or export-control requirements, this sovereign-AI framing has practical purchase implications. A model developed, trained, and licensed under South Korean governance is a structurally different risk profile than one hosted through a US hyperscaler or derived from a Chinese foundation model. This is the same structural argument that drove Microsoft’s recent sovereign AI partnership with Mistral for European regulated sectors, and it applies equally to Asia-Pacific organizations looking for options.
Deployment Roadmap and Ecosystem Expansion
Upstage plans to apply Solar Open 2 to Daum, one of South Korea’s major consumer portals, via a conversational agent interface. Through its “Timely” (also referred to as “Tamnlee”) AI agent platform, the company will support agent adoption across Korean local governments and public institutions. Commercial AX (AI transformation) services are planned across finance, legal, healthcare, defense, manufacturing, consumer goods, robotics, and education verticals.
Upstage also announced plans to list Solar Open 2 on OpenRouter and Hermes Agent, making it accessible via API without self-hosted infrastructure for teams that want to evaluate it before committing to a deployment.
The architecture is designed to support the full stack: open weights for organizations that need full control and customization, API access for teams that want to start fast, and the Timely agent platform for those that want a managed deployment path.
What Enterprise AI Leaders Should Do Now
Solar Open 2 does not require immediate action, but it merits evaluation for specific use cases. Teams building agentic workflows where token cost accumulates across many inference steps should benchmark it against the closed APIs they currently use. Teams operating under data-residency or sovereignty constraints that have ruled out Chinese open-weight alternatives should assess whether Solar Open 2 fits their deployment requirements. Teams planning domain-specific agent fine-tuning should review the Solar License terms alongside other open-weight options.
For teams already operating at scale, the Databricks Unity AI infrastructure framework is one relevant deployment layer for evaluating and serving models like Solar Open 2 alongside other foundation models under unified governance.
The headline number from this release is not 250 billion parameters. It is 15 billion active parameters, 86.8 on Ko-GDPval, and two H200 GPUs. That combination represents what agent-native open-weight model design looks like when it is built for production rather than for benchmark tables.