Korea's four sovereign AI models: open weights, three teams advance
Four open-weight models of 250B parameters or more came out of South Korea's sovereign AI project. Compare their sizes, licenses and hardware needs, and see which teams advanced.
South Korea’s government-backed sovereign AI foundation model project produced four very large open-weight language models in July and August 2026: K-EXAONE 2.0 from LG, A.X K2 from SK Telecom, Motif 3 from Motif Technologies and Solar Open 2 from Upstage. In late August the ministry that runs the project said three of the four teams would go on to the next round and one would not. This piece compares what the four offer, and all of them are on our Korea list.
The four models side by side
| K-EXAONE 2.0 | A.X K2 | Motif 3 | Solar Open 2 | |
|---|---|---|---|---|
| Developer | LG AI Research | SK Telecom | Motif Technologies | Upstage |
| Released | 31 July | 29 July | August | 22 July |
| Parameters (active) | 750B (37B) | 688B (33B) | 314B (13.2B) | 250B (15B) |
| Context | 262K | 128K, 256K with YaRN | 256K | 1M |
| Languages | 10 | 5 | English and Korean | 3 |
| License | Apache 2.0 | Apache 2.0 | MIT | Apache 2.0 plus a naming rule |
| Smallest build listed | NVFP4, 526 GB | NVFP4, 398 GB | NVFP4, 187 GB | INT4 by Nota AI, 143 GB |
All four are text-only mixture-of-experts models, and all four use 8 routed experts per token plus one shared expert, drawn from 256 routed experts at LG and SK Telecom, 320 at Upstage and 384 at Motif.
What the ministry reported
The Ministry of Science and ICT scored the teams out of 100: 40 points for benchmarks, 35 for expert review and 25 for user evaluation, with 49 professional AI users and 185 citizens taking part. The benchmark part included nine Artificial Analysis Intelligence Index tests that Artificial Analysis ran itself. Upstage, SK Telecom and LG advanced “by a narrow margin”. Motif Technologies did not, although the ministry says Motif 3 had the highest Artificial Analysis score of any model from a country other than the US and China. The best Korean score on that index was 47 in August, up from 32 in January, which the ministry ranks sixth among open-weight models. The teams that advanced are to receive about 1,000 B200 GPUs in the second half of 2026, up from about 768. These are the ministry’s own figures.
Licenses
LG moved K-EXAONE from its earlier license to Apache 2.0 for this release, so none of the four has a revenue cap or a territory limit. Motif 3 uses the standard MIT license, and Motif asks, without requiring it, that you mention Built with Motif 3. The one extra condition is Upstage’s Solar License: any AI model you build on Solar Open 2 and distribute must have a name that begins with Solar and show the words Built with Solar.
Hardware and software
None of the four fits on a consumer GPU. Solar Open 2’s card lists 4 H200 GPUs as the minimum, Motif 3’s commands use eight H200 or B200 GPUs, A.X K2’s card recommends at least 800 GiB of GPU memory, and K-EXAONE 2.0’s examples use two nodes of 8 H200 GPUs. All four also need a vendor build rather than stock software: LG’s forked SGLang or vLLM plus a forked Transformers, SK Telecom’s vLLM fork, Upstage’s vLLM image or fork, and Motif’s vLLM fork with trust-remote-code. The VRAM calculator gives a first estimate, and self-hosting explains what running one of these involves.
Which to try first
The score tables come from the developers and use different rivals, so we do not rank the four by them, but they show different strengths. In SK Telecom’s table A.X K2 leads on math and Korean tests, with 97.1 on AIME26 and 80.5 on KMMLU-Pro, yet scores 36.0 on Terminal Bench v2.1 against 61.8 for DeepSeek-V4 Flash and GLM-5.1. Motif reports 74.9 on Terminal-Bench 2.1 but trails on GPQA Diamond. LG says K-EXAONE 2.0 averages 70.1 over 24 benchmarks, and in its table the model leads the three rivals LG chose only on a long-context test and two Korea-focused safety tests. Solar Open 2 has the longest context and the lowest stated GPU minimum. Test two or three on your own Korean tasks, using the guide to evaluating models yourself.
Sources
- Phase 2 Evaluation Results Unveiled for the Sovereign AI Foundation Model Project (Ministry of Science and ICT) msit.go.kr
- K-EXAONE-2.0-750B-A37B model card (Hugging Face) huggingface.co
- A.X K2 model card (Hugging Face) huggingface.co
- Motif 3 model card (Hugging Face) huggingface.co
- Solar-Open2-250B model card (Hugging Face) huggingface.co
- Solar-Open2-250B LICENSE (Upstage Solar License) huggingface.co
Models in this story
| 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 | |
| Solar Open 2 Upstage | 250B 15B active | 1M | Upstage Solar (Conditional) | 157 GB Big-memory desktop |