Upstage Releases Solar Open 2 Agent Model
Upstage released Solar Open 2 on July 22 as a 250-billion-parameter open-weight model that activates 15 billion parameters per token. The company says its hybrid-attention architecture supports a 1 million-token context window and can run on two Nvidia H200 GPUs when quantized; those deployment and benchmark claims still need independent testing on real agent workloads.
Upstage released the weights for Solar Open 2 on July 22, positioning the model for agent workflows such as tool calling, coding and document-heavy office work. The company's model card describes a 250-billion-parameter mixture-of-experts model that activates 15 billion parameters for each token and supports English, Korean and Japanese.
The weights are available on Hugging Face under the Upstage Solar License, which permits commercial use. Independent Korean reporting from Edaily also confirmed the release and its connection to South Korea's sovereign foundation-model program.
A hybrid design for long contexts
Solar Open 2 combines softmax and linear attention across 48 layers. Its repeating pattern uses one softmax-attention layer followed by three linear-attention layers, so only 12 layers retain a conventional KV cache. Upstage says this design supports contexts of up to 1 million tokens while reducing the memory growth associated with an all-softmax stack.
The model has 320 routed experts and one shared expert, selecting eight routed experts plus the shared expert per token. That architecture makes the 15-billion active-parameter figure more useful for estimating inference work than the 250-billion total alone, although memory use also depends on the stored weights, quantization, context length and serving implementation.
Upstage's release page lists four Nvidia H200 GPUs for BF16 inference or two H200s with quantization. The Hugging Face quickstart, however, assumes eight GPUs with at least 141 GB of memory each and warns that actual requirements depend on context length and serving settings. Teams should therefore treat the two-GPU statement as a configuration-specific claim, not a general capacity estimate.
Benchmark claims need workload testing
Upstage reports scores of 80.0 on IFBench, 58.2 on MCP-Atlas and 16.6 on APEX-Agents, alongside strong knowledge, coding and Korean-language results. These are vendor-reported evaluations. The retrieved sources do not provide independent reruns or enough detail to infer production reliability from those numbers alone.
For teams evaluating Solar Open 2, the practical test is whether its architecture produces dependable outcomes under repeated tool calls and long task histories. Useful trials should measure successful task completion, malformed tool calls, recovery after tool errors, latency, peak memory and cost under the team's own context distribution. That evidence matters more than a single benchmark ranking for deciding whether the model fits an agent deployment.
Key Points
- 1Solar Open 2 has 250 billion total parameters, activates 15 billion per token, and is released as open weights under the Upstage Solar License.
- 2Its 48-layer hybrid-attention design uses a one-softmax-to-three-linear pattern and carries a company-stated context limit of 1 million tokens.
- 3Upstage's two-H200 quantized setup and benchmark results are configuration-specific vendor claims that require independent workload testing.
Scoring Rationale
Solar Open 2 is a technically notable open-weight release with sparse activation, a 1 million-token stated context window and explicit agent-workflow positioning. Its relevance is high for teams evaluating self-hosted long-context agents, but the deployment and benchmark claims remain vendor-reported and need independent validation.
Sources
Primary source and supporting public references used for this report.
Practice with real Ride-Hailing data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ride-Hailing problems
