LG AI Research releases K-EXAONE 2.0
The open-weight multilingual MoE model has 750B total parameters and a 256K-token context window.
Why it matters
K-EXAONE 2.0 adds another large Apache 2.0 open-weight model to the multilingual model landscape, with explicit emphasis on Korean sociocultural safety. Its long-context and agentic coding gains may make it relevant for developers evaluating alternatives to other open-weight systems.
The key points
- 1.750B-parameter MoE activates about 37B parameters per token.
- 2.Context length increases to 256K tokens.
- 3.Language coverage expands from six to ten languages.
LG AI Research released K-EXAONE 2.0, an open-weight multilingual foundation model developed by upcycling and expanding K-EXAONE rather than training from scratch. The Mixture-of-Experts model has 750B total parameters, about 37B activated per token, supports context lengths up to 256K tokens, and expands language coverage from six to ten languages. The report says it improves over K-EXAONE across nine evaluation categories, with the largest gains in agentic coding and long-context understanding.
⚡ Try this today
Read the technical report and evaluate K-EXAONE 2.0 on your own multilingual, long-context, and coding tasks before deployment.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
Enjoyed this brief? Get the next one in your inbox.
More in Models
Qwen3.8 27B scores 52 on Artificial Analysis
Reddit and Hacker News posts highlighted a reported jump in Qwen3.8 27B benchmark results.
Google introduces Gemini 3.7 Flash
Google and DeepMind published posts introducing the Gemini 3.7 Flash model.
Meta releases open-weight Glimmer AI model
The downloadable model contrasts with Meta’s more powerful API-only Muse Spark.