Solar Open2 250B
Solar Open2 250B by Upstage: 250B-A15B hybrid-attention MoE for agentic workflows. 15B active per token, 1M context with NoPE, English/Korean/Japanese.
Model Overview
Capabilities, design details, and architectural traits
Solar Open2 250B - Hybrid-Attention MoE for Agentic Workflows
Solar Open2 250B is Upstage's open-weight large language model purpose-built for agentic use cases including office productivity, document-intensive work, and coding. Its defining architectural choice is a Hybrid-Attention Mixture-of-Experts design that interleaves three linear-attention layers with one softmax-attention layer in a repeating [Softmax, Linear×3] × 12 pattern across 48 layers.
| Trait | Detail |
|---|---|
| Sparse activation | 250B total parameters but only 15B activated per token via top-8 routing among 321 experts (320 routed + 1 shared) |
| NoPE architecture | Linear-attention layers encode token order intrinsically, so positional encoding is removed entirely, lifting the RoPE extrapolation limit to a 1M-token context |
| Minimal KV cache | Only 12 of 48 layers retain a KV cache, holding long-context memory to roughly a quarter of an all-softmax model of the same shape |
| Selective weight transfer | Initialized from Solar Open 1 (102B) by carrying over only 2.3% of weights that survive the architectural change; everything else is randomly initialized |
| Agentic specialization | Designed for tool calling, multi-step reasoning, and end-to-end task execution |
| Multilingual scope | Trained and evaluated in English, Korean, and Japanese |
The hybrid attention stack is central to the model's identity. By replacing three out of every four attention layers with linear attention, Solar Open2 250B achieves large-model capacity at a fraction of the inference cost. The removal of rotary positional encoding entirely (NoPE) is a direct consequence of the linear-attention design, enabling the 1M-token context window without the extrapolation limits typical of RoPE-based models.
The selective weight transfer from Solar Open 1 is another distinctive trait: rather than training from scratch, Upstage reused the small subset of weights compatible with the new architecture to raise the starting point and accelerate early convergence at 250B scale.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Upstage | Anthropic | Anthropic |
| Release Date | August 12, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 1M | 1M | 1M |
| Pricing (per 1M tokens) | |||
| Input Pricing | Free Best Input Pricing | $4 | $2 |
| Output Pricing | Free Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | textfile | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 24.7 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 45.0 | - | - |
| Agentic Index | 27.8 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from Upstage
Other models by Upstage
Top AI Models
Leading alternatives by intelligence score
