K-EXAONE 2.0
K-EXAONE 2.0 is LG AI Research's open-weight mixture-of-experts language-model release. The flagship checkpoint is K-EXAONE-2.0-750B-A37B, with roughly 750 billion total parameters, 37 billion active parameters, ten listed languages, and Apache 2.0 licensing.
The release combines unusually large open weights with an activated footprint far below the total parameter count and provides official quantized variants. That expands the deployment choices for teams that can operate cluster-scale infrastructure, but it does not make the model a desktop-class local model.
LG AI Research's current Hugging Face repository is public and ungated, publishes BF16 weights, an Apache 2.0 license, a technical report, ten language tags, a 256-expert configuration that selects eight experts per token, and official FP8, NVFP4, and DSpark variants. The 750B-A37B identity comes from the official repository name and published artifacts; community discussion is used only for access, size, and deployment questions. Launch benchmark comparisons remain vendor-reported.
- Evaluate large multilingual MoE behavior on controlled internal tasks.
- Test Korean and English instruction following and reasoning.
- Compare official full-precision and quantized deployment variants.
- Study tool-call formatting through the published chat template.
The primary repository is LGAI-EXAONE/K-EXAONE-2.0-750B-A37B. The current Hugging Face API shows public, ungated Transformers-compatible weights and approximately 749.4 billion stored parameters, consistent with the 750B family label.
The official metadata declares Apache 2.0, which is materially different from a custom research-only license. Even so, the full BF16 checkpoint is cluster-scale. Operators should choose among the full, FP8, NVFP4, and DSpark artifacts based on verified runtime support instead of assuming that 37B active parameters means 37B storage or memory.
LG AI Research includes evaluation results in the model-card and report bundle. Treat those as vendor-reported until the same checkpoints, prompts, reasoning settings, sampling parameters, serving stack, and judge conditions are independently reproduced.
K-EXAONE 2.0 FAQ
Common questions about K-EXAONE 2.0.
Is K-EXAONE 2.0 open source?+
The flagship model publishes open weights under Apache 2.0, and its Hugging Face repository is currently public and ungated. Review the repository notice and every dependency before treating the whole serving stack as having one license.
Can K-EXAONE 2.0 run like a 37B model?+
No. A37B describes parameters activated during a token pass, not the checkpoint size. The model still contains roughly 750B total parameters, so storage, memory, networking, and expert-parallel serving remain cluster-scale concerns.
Does K-EXAONE 2.0 have a public hosted API?+
The verified release evidence is for downloadable model artifacts, not a stable first-party API model ID with public pricing and rate limits. Treat hosted endpoints from third parties as separate provider contracts.