News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnK-EXAONE 2.0

K-EXAONE 2.0

Open-weight multilingual language model by LG AI Research

LG AI Research released K-EXAONE 2.0, a large-scale multilingual model with 750 billion parameters. It uses a mixture of experts design and supports ten languages under an open licence.

Share on WhatsApp

  • 750 billion total parameters with 37 billion active
  • Supports ten languages including Korean and English
  • Released under Apache License 2.0
  • Uses mixture of experts with 78 main layers
  • Inference sped up 3–5× with speculative decoding

What it is

K-EXAONE 2.0 is a frontier-scale language model built by LG AI Research. It contains 750 billion parameters in total, with 37 billion active for each token processed. The system expands on the earlier K-EXAONE model through upcycling, which increases both depth and width. It supports ten languages and offers strong performance in long-context retrieval and safety tasks. The model is designed for agentic workflows and complex reasoning.

How it came about

The developers scaled the previous K-EXAONE model to more than three times its original size. This process involved upcycling followed by continual pretraining and difficulty-focused mid-training. Post-training steps were also applied to refine the model. The team aimed to create a foundation model with high-level intelligence suitable for modern agentic AI applications.

Inside the model

The architecture uses a mixture of experts design with 78 main layers, including two dense heading layers and 76 sparse layers. It employs 19 blocks of sliding window attention and global attention. To handle stability in deeper layers, the team clamped activations after two SwiGLU branches. Training included careful calibration of data recipes to improve agentic coding and long-context understanding.

What it can do

The model shows substantial improvements over its predecessor in long-context retrieval and safety. It supports advanced reasoning and agentic intelligence, allowing it to manage complex workflows. Performance is consistent in agentic coding tasks. The system can operate in reasoning mode for high accuracy or non-reasoning mode for lower latency. It also supports tool calling for agent integration.

How to use it

LG AI Research releases the model weights under the Apache License 2.0. This permits inspection, deployment, and building upon the model by the broader community. Users can run it via SGLang or vLLM forks provided by the team. Inference can be accelerated using MTP or DSpark speculative decoding methods, which speed up generation by three to five times.

Safety and testing

The company states that efforts were made to exclude harmful and biased information from training data. However, it acknowledges that inappropriate responses may still occur due to statistical patterns in the data. Biased outputs related to age, gender, or race are possible. Users are prohibited from engaging in malicious activities that violate ethical principles. The model does not reflect the views of LG AI Research.

Questions people ask about K-EXAONE 2.0

How many parameters does K-EXAONE 2.0 have?

The model has 750 billion total parameters. It uses a mixture of experts architecture, so only 37 billion parameters are active for each token processed during inference.

What languages does it support?

It supports ten languages: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish, and Portuguese. This is an expansion from the six languages supported by its predecessor.

Is the model open source?

Yes, the model weights are released under the Apache License 2.0. This allows the broader AI ecosystem to inspect, deploy, and build upon the model freely.

How can I speed up inference?

The model supports two speculative decoding methods called MTP and DSpark. These methods can accelerate model generation by approximately three to five times, reducing latency for long tasks.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on K-EXAONE 2.0 yet, so the facts on this page come from the sources listed above: huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.