What is K-EXAONE 2.0?
K-EXAONE 2.0 is LG AI Research's flagship language model, released with open weights on Hugging Face on 29 July 2026 and the largest model built in Korea. It is a mixture-of-experts model with 750 billion parameters, 37 billion active per token, made by enlarging the earlier 236B K-EXAONE and training it further. It has a 262,144-token context, reasons step by step, and covers ten languages: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish and Portuguese.
Unlike earlier EXAONE models, which limited commercial use, it is under the Apache 2.0 licence. LG ships BF16, FP8 and NVFP4 weights plus a DSpark speculative-decoding variant, and says multi-token prediction and DSpark speed up generation by about 3 to 5 times. On Hugging Face it has 159 likes and 14,254 downloads in the last 30 days.
LG publishes no per-token price. FriendliAI offers it as a dedicated endpoint, and LG runs a demo. Its serving examples use two nodes of eight NVIDIA H200 GPUs.
How much does K-EXAONE 2.0 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| FriendliAI dedicated endpoint | GPU-time billing | No per-token rate published |
Free weights under Apache 2.0. No per-token price published; FriendliAI offers dedicated endpoints billed for GPU time.
Is K-EXAONE 2.0 worth it?
K-EXAONE 2.0 is worth considering for two things: Korean-language work and finding facts in very long documents. LG reports 94.4 on OpenAI-MRCR, a test of retrieving details buried in long context, against 92.9 for DeepSeek-V4-Pro and 71.5 for GLM-5.1, and 89.6 on Ko-LongBench. Its safety scores are the highest in LG's table, 99.8 on KGC-Safety and 89.5 on ROK-Fortress, though LG built KGC-Safety itself.
On almost everything else it is behind. In the same table it scores 43.8 on Terminal-Bench 2.1 against 64.0 for DeepSeek-V4-Pro and 61.8 for GLM-5.1, and 18.3 on Humanity's Last Exam against 37.7. Even on the Korean KMMLU-Pro test it trails, 69.1 against 80.5. Artificial Analysis's v5.0.1 index scores it 8, against a median of 23 for comparable open models.
The catch: running it is a cluster project. The quickstart needs LG's own forks of SGLang and Transformers, serves across two eight-GPU H200 nodes, and LG warns of output collapse on B200 GPUs unless a flag is set. There is no public per-token API, so you cannot try it cheaply before buying the hardware.
Our take: K-EXAONE 2.0 makes sense for a Korean public body or bank that needs an Apache-licensed model from a domestic lab and works mostly with long Korean documents. Anyone choosing purely on capability gets more from DeepSeek-V4-Pro or GLM-5.3.
Where it earns its place
A Korean bank's compliance team
Search long regulatory filings and contracts in Korean on its own servers
Self-host the FP8 weights on two H200 nodes and use the 262,144-token context for whole-document questions. Apache 2.0 means no licence negotiation, and LG's 89.6 on Ko-LongBench suits this work.
A public-sector AI team with a sovereign-model mandate
Deploy a domestically built frontier model with strict safety behaviour
K-EXAONE 2.0 leads LG's table on KGC-Safety and ROK-Fortress. Budget for a dedicated cluster or a FriendliAI dedicated endpoint, since no per-token service exists.
Before you commit
- Setup
- Install LG's SGLang and Transformers forks, then serve with tensor parallelism across two nodes of eight H200 GPUs; vLLM recipes are also on the card, but DSpark decoding is not supported on vLLM.
- Integrations
- OpenAI-compatible endpoint once served; Qwen3 reasoning and tool-call parsers; FriendliAI dedicated endpoints.
- Data and privacy
- Apache 2.0 weights, so self-hosting keeps data in-house. LG's knowledge cutoff is the second quarter of 2025.
- Cost at scale
- No per-token price. Self-hosting needs sixteen H200 GPUs for the reference setup; FriendliAI bills dedicated endpoints for GPU time.
- Limits
- 262,144-token context. On B200 GPUs LG says to add --disable-prefill-cuda-graph to avoid generation collapse.
What’s good
- Apache 2.0 licence, a first for LG's large models
- 94.4 on OpenAI-MRCR long-context retrieval, per LG
- Top safety scores in LG's comparison
- Ten languages including Korean, Vietnamese and Polish
What’s not
- Trails DeepSeek-V4-Pro and GLM-5.1 on coding, maths and knowledge in LG's own table
- Scores 8 on Artificial Analysis's index against a peer median of 23
- Needs two eight-GPU H200 nodes and forked serving code
- No public per-token API
Best for
Korean organisations that need a permissively licensed domestic model for long Korean documents and can host it themselves.
K-EXAONE 2.0 alternatives
K-EXAONE 2.0: common questions
- Is K-EXAONE 2.0 worth it?
- A properly open Korean frontier model with excellent long-document recall, but by LG's own tables it trails the Chinese open models on most work. We rate it 3.4 out of 5.
- How much does K-EXAONE 2.0 cost?
- Free weights under Apache 2.0. No per-token price published; FriendliAI offers dedicated endpoints billed for GPU time.
- What are the downsides of K-EXAONE 2.0?
- Trails DeepSeek-V4-Pro and GLM-5.1 on coding, maths and knowledge in LG's own table; scores 8 on Artificial Analysis's index against a peer median of 23; needs two eight-GPU H200 nodes and forked serving code; no public per-token API.
- Who is K-EXAONE 2.0 best for?
- Korean organisations that need a permissively licensed domestic model for long Korean documents and can host it themselves.
- How much does K-EXAONE 2.0 cost for a team or in production?
- No per-token price. Self-hosting needs sixteen H200 GPUs for the reference setup; FriendliAI bills dedicated endpoints for GPU time.
- Is K-EXAONE 2.0 safe for company data?
- Apache 2.0 weights, so self-hosting keeps data in-house. LG's knowledge cutoff is the second quarter of 2025.
- What are the best K-EXAONE 2.0 alternatives?
- The closest alternatives we have reviewed are DeepSeek-V4-Pro, GLM-5.3, Solar Open 2 and Mistral Large 4.
Answers checked Nov 7, 2026.