Krafton unveiled a preview version of its self-developed 30B-class large language model, Raon-1 Mini, on Oct. 21. The model supports Korean and English and was released on Hugging Face ahead of an official launch under an Apache 2.0 license. Krafton said the model uses a mixture-of-experts structure and its own linear attention technology to improve response speed and inference cost efficiency.

Google has unveiled Deep Review, an AI tool designed to find logical contradictions and mathematical errors in research papers and technical documents. The system uses multiple AI agents to review sections in parallel and then consolidates results while cross-checking external sources to reduce duplicate or incorrect findings. Google said tests on 26 math and computer science papers found it detected 89.7 percent of 29 confirmed errors, outperforming Gemini 3.1 Pro used alone.