Solar Open 2

Upstage's 250B mixture of experts with 15B active and a 1M context for English, Korean and Japanese. Apache 2.0 plus a naming rule.

In short. Solar Open 2 is a large AI model from the South Korean company Upstage, built for office work, document-heavy tasks and coding agents. It works in English, Korean and Japanese and can read up to a million tokens. Companies can use it commercially, but any model you build from it and share must have a name that starts with Solar.

Summary

Upstage released Solar Open 2 on 22 July 2026, with a technical report the same day. It is a mixture-of-experts model with 250 billion parameters and 15 billion active per token, built for agentic work such as office productivity, document-heavy tasks and coding, and developed under South Korea's government-backed sovereign AI foundation model project. It has 48 layers, and its attention alternates three linear-attention layers with one softmax layer and uses no positional encoding, so only 12 of the 48 layers keep a KV cache and the context window reaches 1M tokens. There are 320 routed experts plus one shared expert, with 8 routed experts used per token. Upstage trained on about 12 trillion tokens, started from its 102B Solar Open 1 by transferring only the roughly 2.3% of weights that survive the architecture change, and supports English, Korean and Japanese.

In Upstage's tables Solar Open 2 scores 86.2 on MMLU-Pro, 70.4 on SWE-Bench Verified and 95.7 on AIME 2026, against 85.9, 73.8 and 97.0 for DeepSeek-V4-Flash (max) and 84.6, 73.0 and 92.3 for MiMo-V2.5. On Korean tests it scores 90.7 on CLIcK, ahead of those two models and of GPT-5.4 mini. Upstage marks some Korean tests as in-house benchmarks. All numbers are Upstage's own.

The license is the Upstage Solar License: Apache 2.0 with an added rule that any AI model you create, train, fine-tune or improve with Solar Open 2 and then distribute must have a name that begins with Solar and show the words Built with Solar. Nothing else limits commercial use. The BF16 weights are about 501 GB, and the card lists 4 NVIDIA H200 GPUs as the minimum and 8 as recommended. For smaller setups it points to quantized builds published by Nota AI in NVFP4 (about 153 GB) and INT4 (about 143 GB). The card recommends vLLM through Upstage's Docker image or fork, and Upstage's Transformers branch for experiments.

Good for

  • Agentic work in English, Korean and Japanese, such as office productivity, document-heavy tasks and coding
  • Long-context workloads, since only 12 of the 48 layers keep a KV cache
  • Korean-language products, where Upstage's own tests show strong Korean knowledge and office-task results
  • Companies that want a commercial-friendly license and can accept the naming rule for derived models

Watch out for

  • Any derivative AI model you distribute must have a name that starts with Solar and show Built with Solar
  • The BF16 weights are about 501 GB, and the card lists 4 H200 GPUs as the minimum
  • It needs Upstage's vLLM image or fork, or its Transformers branch, rather than stock releases
  • Only English, Korean and Japanese are supported
  • All benchmark numbers are Upstage's own, and several Korean tests are in-house

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Solar Open 2 Nota INT4 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Solar Open 2 (BF16) BF16 250B 15B active 575 GB 288 GB 173 GB Big-memory desktop
Solar Open 2 Nota NVFP4 NVFP4 250B 15B active 575 GB 288 GB 168 GB Big-memory desktop
Solar Open 2 Nota INT4 INT4 250B 15B active 575 GB 288 GB 157 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

Apache 2.0 with one added rule. You can use the model commercially and build on it freely, but an AI model you create from it and distribute must have a name that starts with 'Solar' and must show the words 'Built with Solar'.

  • Derived models must carry the original name
  • Attribution or model name must be shown

More about Upstage Solar

See also

Frequently asked questions

What is Solar Open 2?
Solar Open 2 is a large AI model from the South Korean company Upstage, built for office work, document-heavy tasks and coding agents. It works in English, Korean and Japanese and can read up to a million tokens. Companies can use it commercially, but any model you build from it and share must have a name that starts with Solar.
Can I use Solar Open 2 commercially?
Yes, with conditions. Solar Open 2 is released under the Upstage Solar, which allows commercial use but adds extra rules: Derived models must carry the original name; Attribution or model name must be shown. Read the license text before you ship a product.
Is Solar Open 2 open source?
Solar Open 2 is open-weight, not open source in the strict sense. You can download the model files, but the Upstage Solar adds conditions that the Open Source Initiative definition does not allow.
Can I run Solar Open 2 on my own computer?
Only on a machine with a lot of memory. The smallest variant, Solar Open 2 Nota INT4, needs about 157 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was Solar Open 2 released?
Solar Open 2 was released by Upstage on July 22, 2026.
How long a context window does Solar Open 2 have?
Solar Open 2 supports a context window of up to 1M tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
K-EXAONE 2.0 LG AI Research 750B 37B active 256K Apache 2.0 (Permissive) 518 GB GPU server
A.X K2 SK Telecom 688B 33B active 256K Apache 2.0 (Permissive) 438 GB Big-memory desktop
Motif 3 Motif Technologies 314B 13.2B active 256K MIT (Permissive) 206 GB Big-memory desktop
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone