Upstage Releases Solar Open 2 Agent Model
For practitioners, open-weight MoE models with long context create a distinct evaluation target for agent systems: total parameters matter less than active parameters, memory behavior, tool-use reliability, and deployment hardware. Upstage released the weights for Solar Open 2, a 250B-A15B hybrid-attention MoE model, on Hugging Face. The model card lists a 1 million-token context window, 15B active parameters per token, and support for English, Korean, and Japanese. Upstage reports benchmark results against several open models, but the available sources do not provide independent benchmark validation. Chosun reports that quantized deployments can run on two Nvidia H200 GPUs and that Upstage intends to apply the model to Daum and its Timely AI agent platform.
An architecture aimed at agent economics
For practitioners, agent deployments often expose a gap between a model's headline parameter count and its operational cost. Long task trajectories can multiply inference calls, tool invocations, and context retention requirements. Hybrid-attention MoE designs are therefore worth evaluating on active parameters, KV-cache requirements, latency under repeated calls, and tool-use reliability rather than on aggregate parameter count alone.
Upstage released the weights for Solar Open 2 on Hugging Face. According to the model card, Solar Open 2 is a 250B-A15B model: it contains about 250.3 billion total parameters and activates 15 billion parameters per token. Chosun and Digital Today report that the release was developed through South Korea's government sovereign AI foundation model project, where Upstage is the lead organization.
The Hugging Face model card describes Solar Open 2 as an open-weight model for agentic workflows, including tool calling, multi-step reasoning, document work, office productivity, and coding. It supports English, Korean, and Japanese, has a stated context length of 1 million tokens, and is distributed under the Upstage Solar License.
Hybrid attention and memory behavior
According to the model card, the model uses 48 layers with a repeating pattern of one softmax-attention layer followed by three linear-attention layers. It uses 321 experts, comprising 320 routed experts and one shared expert; routing activates eight routed experts plus the shared expert for each token.
Upstage's documentation states that the linear-attention layers maintain token order in recurrent state, allowing the model to use NoPE, or no positional encoding, instead of rotary positional embeddings. The documentation also states that only 12 of its 48 layers retain a KV cache, which it characterizes as roughly one quarter of the long-context memory of an all-softmax model with the same shape.
Industry context
linear attention and sparse expert routing can reduce some memory and compute burdens, but production behavior still depends on kernel maturity, batch size, quantization method, context distribution, and the agent framework's tool-call loop. Teams testing this model should separately measure prefill throughput, decode throughput, peak memory, and accuracy on their own multistep tasks.
The model card reports approximately 12 trillion pre-training tokens, training on Nvidia B200 GPUs, and 2 million GPU-hours of training. It also states that Solar Open 2 was initialized through selective transfer from Solar Open 1, carrying forward 2.3% of weights that survived the architecture change.
Performance claims require local validation
Upstage reports that Solar Open 2 is competitive with leading open-weight models on agent benchmarks. Digital Today reports that Upstage compared the model with DeepSeek V4 Flash, Mistral Medium 3.5, and Cohere Command A+ across MMLU-Pro, LiveCodeBench, IFBench, MCP-Atlas, and APEX-Agents. Chosun specifically reports Upstage's claims of stronger results than DeepSeek V4 Flash and Mistral Medium 3.5 on IFBench and MCP-Atlas.
According to Chosun and Digital Today, Solar Open 2 scored 86.75 on Ko-GDPval, a Korean industrial-task benchmark. The outlets report Upstage's comparison of that result with DeepSeek V4 Pro, described as more than six times larger by parameter scale. Digital Today also reports Upstage's claim that the model exceeded GPT-5.4 Mini and Claude Haiku 4.5 on an average of Korean-language benchmarks.
For practitioners, vendor benchmark comparisons are useful screening signals rather than deployment evidence. The retrieved sources do not provide independent benchmark runs, prompt configurations, sampling settings, tool schemas, or task-level failure analysis. Agent evaluations should include malformed tool calls, recovery after tool errors, long-horizon state consistency, multilingual retrieval quality, and cost per successfully completed task.
Deployment claims and reported applications
Chosun reports that a quantized Solar Open 2 deployment can run on two Nvidia H200 GPUs. The model card itself describes the architecture as designed for efficient inference in long-context settings, but hardware feasibility does not establish a production latency or concurrency profile. Quantization choices and serving-stack support for hybrid attention and MoE routing remain material implementation variables.
Chosun reports that Upstage intends to use Solar Open 2 in Daum, expanding from search and summarization to conversational agent services. The same report states that the company intends to support agent adoption by local governments and public institutions through Timely, its AI agent platform, and to pursue industry-specific AI transformation services in finance, law, and healthcare alongside domestic NPUs.
Observed patterns in similar transitions: releasing weights gives engineering teams the option to inspect deployment constraints and run domain-specific evaluations, while the real differentiators for enterprise agents tend to be integration quality, governance, observability, and evaluation coverage. Solar Open 2's combination of sparse activation, hybrid attention, and Korean-language positioning makes it a notable candidate for teams whose workloads combine long documents, repeated tool use, and Korean or Japanese language requirements.
Key Points
- 1Solar Open 2 combines 250B total parameters with 15B active parameters, making active compute and memory central evaluation metrics.
- 2Its hybrid attention design targets 1 million-token contexts with fewer KV-cache layers, requiring serving-stack validation before production adoption.
- 3Upstage reports strong agent and Korean benchmark results, while practitioners need independent task-level testing for tool-use reliability.
Scoring Rationale
This is a notable open-weight model release with a large sparse MoE architecture, 1 million-token context, and explicit agent-workflow targeting. Its hybrid-attention design and Korean-language emphasis are technically relevant to teams evaluating self-hosted long-context agents, although reported benchmark results require independent validation.
Sources
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
