South Korea’s Trilion Labs has made history by open sourcing the full checkpoints of its 70B-parameter AI model Tri-70B, enabling researchers and developers worldwide to build and customise large-scale AI efficiently.
South Korean AI startup Trilion Labs has become the first company globally to open-source the full intermediate checkpoints of a 70-billion-parameter AI model, Tri-70B, the largest LLM developed from scratch in the country. The model and its checkpoints are released under the Apache 2.0 license, enabling both research and commercial use, with previous Tri series models (0.5B, 1.8B, 7B) also available under the same licence.
Unlike typical model releases, Trilion Labs’ checkpoints include step-by-step outputs from the training process, offering a blueprint of the AI’s development. This transparency allows developers to customise models efficiently without training from scratch, saving time and resources. A company representative said, “Beyond simply disclosing technology, we aim to establish a foundation that allows researchers, startups, students, and developers to collaboratively experiment and create derivative models.”
The open source move has generated significant attention. A post on the Local Llama Reddit community ranked first in popularity, receiving enthusiastic responses and direct inquiries, reflecting a strong global developer interest.
Founded in 2024 by Shin Jae-min, who previously contributed to HyperCLOVA X at Naver, Trilion Labs secured approximately 9 billion KRW in funding and released Tri-70B within a year of establishment.
The move underscores a growing global shift toward open source AI, with major players such as OpenAI, Meta, and Google increasingly sharing models or components. In South Korea, companies including LG, Naver, and Kakao, as well as innovators like Dennis Hong in robotics, are following similar paths. Trilion Labs’ open source release positions South Korea as a key contributor to the global AI ecosystem, enabling collaborative development and wider commercial application of large-scale AI models.
