Solar Open 2 vs Motif 3: which Korean agent model should you run?
Solar Open 2 has 250B parameters, a 1M context and a naming rule; Motif 3 has 314B, a 256K context and the plain MIT license. Compare hardware, languages and licenses.
Verdict. Choose by license and context first. Solar Open 2 has 250B parameters with 15B active, reads up to 1M tokens, supports English, Korean and Japanese, and lists 4 H200 GPUs as its minimum, but its Apache 2.0 license adds a rule that any model you build on it and distribute must have a name starting with Solar. Motif 3 has 314B parameters with 13.2B active, uses the standard MIT license with no naming rule, has a 256K context and names English and Korean as its languages, and its commands use eight H200 or B200 GPUs. Motif reports 74.9 on Terminal-Bench 2.1, the strongest agent result in its own table, but trails on GPQA Diamond and math. Upstage reports 86.2 on MMLU-Pro and 70.4 on SWE-Bench Verified. All of these numbers are the developers' own.
Specifications
| Specifications | Solar Open 2 | Motif 3 |
|---|---|---|
| Developer | Upstage | Motif Technologies |
| Released | July 22, 2026 | August 2026 |
| Total parameters | 250B (Solar Open 2 (BF16)) | 314B (Motif 3 (BF16)) |
| Active parameters | 15B | 13.2B |
| Architecture | Mixture of experts | Mixture of experts |
| Context window | 1M tokens | 256K tokens |
| Input | Text | Text |
| License | Upstage Solar (Conditional) | MIT (Permissive) |
| Memory for the smallest variant (4-bit) | 157 GB (Solar Open 2 Nota INT4) | 206 GB (Motif 3 NVFP4) |
Performance claims come from the developer and have not been independently verified.
Pick Solar Open 2 if
- You work with very long documents, since it reads up to 1M tokens.
- You need Japanese alongside English and Korean.
- You have a smaller GPU budget: the card lists 4 H200 GPUs as the minimum, and Nota AI publishes quantized builds of 143 to 153 GB.
- You can accept the naming rule for any derivative model you distribute.
Pick Motif 3 if
- You want the plain MIT license, with no naming rule and no required credit.
- Your work is terminal-based coding agents, where Motif reports 74.9 on Terminal-Bench 2.1.
- You want to study or reuse Motif's own architecture, since it publishes an example repository with its architecture and distributed training framework.
- English and Korean are enough, and a 256K context covers your inputs.