Solar Open2 250B
#13 en Modèles de langage open sourceUpstage · v2 · 250b · depuis 23. Juli 2026 · 7× · vu le 24 juil. 2026
Solar Open 2 250B is an open-weight language model by Upstage with 250 billion total parameters, of which only 15 billion are activated per token thanks to its Mixture-of-Experts architecture. It uses a hybrid-attention design (linear attention combined with softmax attention, without positional encoding) that enables a context window of up to 1 million tokens with efficient inference. The model was developed as part of Korea's government-backed "Sovereign AI" initiative and is optimized for agentic tasks such as tool use, multi-step reasoning, and coding. It officially supports Korean, English, and Japanese, and is available under the commercially usable Upstage Solar License.
Fonctionnalités
| Key Benchmark (%) | 70.4% on SWE-bench Verified |
| Context Window (Tokens) | 1,000,000 tokens |
| License | Upstage Solar License (Apache 2.0-based, permits commercial use) |
| Multimodality | No, text-only language model (no multimodal input); multilingual: Korean, English, Japanese |
| Platform | Hugging Face (model weights), local deployment via vLLM/Transformers, runs on 2x NVIDIA H200 GPUs with quantization |
| Release Date | July 23, 2026 |