News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnSolar Open 2

Solar Open 2

250B parameter open-weight model for agentic tasks

Upstage released Solar Open 2 in 2026 as a large language model designed for complex agent workflows. It uses a hybrid architecture to handle long contexts efficiently while supporting English, Korean and Japanese.

Share on WhatsApp

  • 250B total parameters with 15B active per token
  • Supports 1 million token context windows
  • Trained on 12 trillion tokens
  • Specialised for agentic workflows and tool use
  • Hybrid attention reduces memory for long contexts

What it is

Solar Open 2 is a large language model developed by Upstage. It contains 250 billion parameters but activates only 15 billion during inference. The system targets agentic use cases such as coding and document processing. It supports English, Korean and Japanese. The model handles context windows up to one million tokens. Upstage designed it to balance high capacity with low inference costs through its specific architectural choices.

How it came about

The model builds upon Solar Open 1, which had 102 billion parameters. Developers used selective weight transfer to initialise the new system. Only 2.3 per cent of weights survived the architectural changes. The rest were randomly initialised to accelerate convergence. Training consumed two million GPU hours on NVIDIA B200 hardware. The company released the technical report and model weights in 2026.

Inside the model

The architecture combines softmax and linear attention layers in a repeating pattern. Twelve of the forty-eight layers retain a key-value cache. This design removes the need for rotary positional encoding. The mixture of experts system includes 321 experts, activating eight routed and one shared expert per token. Linear attention layers encode token order in their recurrent state. This approach reduces memory usage for long sequences compared to standard softmax models.

What it can do

The model performs strongly on reasoning and coding benchmarks. It scores 86.2 on MMLU-Pro and 92.4 on LiveCodeBench v6. In Korean evaluations, it achieves 78.4 on KMMLU-Pro and 90.7 on CLIcK. The system excels in agentic tasks, scoring 16.6 on APEX-Agents. It supports tool calling and multi-step reasoning. Performance remains competitive with other large open-weight models in document-intensive work.

How to use it

Upstage provides open weights under the Upstage Solar License. Users can run the model locally using Transformers or vLLM. The recommended setup requires eight NVIDIA H200 or B200 GPUs. Quantized versions are available from NotaAI for smaller hardware configurations. The company offers a custom fork of vLLM to support the unique attention mechanisms and reasoning parsers required for optimal performance.

Questions people ask about Solar Open 2

How many parameters does Solar Open 2 have?

The model has 250 billion total parameters. However, it uses a mixture of experts architecture that activates only 15 billion parameters for each token processed. This allows it to maintain high capacity while keeping inference costs lower than dense models of similar size.

What languages does Solar Open 2 support?

The model supports three languages: English, Korean and Japanese. It shows strong performance in Korean benchmarks such as KMMLU-Pro and CLIcK, reflecting its development background. It is designed to handle multilingual tasks within these specific language families.

What hardware is needed to run Solar Open 2?

Upstage recommends eight NVIDIA H200 or B200 GPUs with at least 141 GB of memory each. Quantized versions are available for smaller setups. The model can also run on four H200 GPUs as a minimum requirement, though performance may vary based on context length and serving settings.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Solar Open 2 yet, so the facts on this page come from the sources listed above: huggingface.co/upstage/Solar-Open2-250B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.