Trillion Labs, a startup focused on artificial intelligence (AI) models, has introduced the most extensive language model to date.
Embracing open source with minimal post-learning…Enhance the use of third-party resources and transparently disclose intermediate checkpoints “for the first time in Korea.”
![Trillion Labs has announced an "open source" initiative, rolling out its model lineup for commercial applications, including the largest Large Language Model (LLM) with 70 billion parameters. [Source = Trillion Labs]](https://wimg.mk.co.kr/news/cms/202509/10/news-p.v1.20250910.05e37acb5d2d41e5937508bd5200ec73_P1.png)
Trillion Labs, an AI model startup, has introduced the most extensive language model (LLM) ever created. It is the leading Korean model available for open source and is notable for its transparent disclosure of the learning process.
On the 10th, Trillion Labs unveiled its proprietary “Tri-70B” model featuring an impressive 70 billion parameters from the pre-learning phase.
The Tri-70B model marks the largest release by Trillion Labs. Parameters function like synapses in the human brain within the AI model, and a larger size equates to enhanced computational capabilities.
This model is pre-trained on approximately 1.5 trillion tokens, with a context window—an input data unit processed simultaneously by AI models—of about 32,000 tokens, specializing in English, Korean, and Japanese.
“We presented it as a base model with minimal post-learning,” stated Trillion Labs. “This is anticipated to aid both academia and industry, allowing researchers and companies easy customization.”
They also revealed a real-time search-specific model that incorporates the latest information alongside AI models geared towards multilingual translation and search functionality.
With the introduction of the Tri-70B model, Trillion Labs proclaimed an “Open Source Month,” releasing all previously available model lineups under Apache 2.0 licenses for commercial use.
Moreover, they fully disclose not only the final model but also the intermediate checkpoints generated throughout training. The aim is to stress model transparency and assist other researchers in studying the model by revealing the entire training process.
This enables both academia and industry to thoroughly investigate the learning processes of large models and to undertake efficient re-learning and applied research.
Trillion Labs highlighted that the release of intermediate checkpoints has been limited by non-profit organizations like Allen Research Institute and Hugging Face, making this a first in Korea.
Shin Jae-min, CEO of Trillion Labs, remarked, “We aspire to create language models not just for their own sake but to disclose all training processes and core techniques to foster transparency in research, laying the groundwork for the development of the global AI research ecosystem, including in Korea.”
