Building world-class language models
Building world-class language models
Building world-class language models
We believe that the future of AI is efficient, deliberate, and responsibly deployed.
We believe that the future of AI is efficient, deliberate, and responsibly deployed.
We believe that the future of AI is efficient, deliberate, and responsibly deployed.
Establishing a New Pareto Frontier
Establishing a New Pareto Frontier
Establishing a New Pareto Frontier
We push the performance–cost frontier of LLMs. Our models deliver state-of-the-art performance per FLOP, achieving competitive results at a fraction of the training cost. By optimizing architecture, data strategy, and training efficiency end-to-end, we prove that frontier-level capability does not require frontier-level spend.
We push the performance–cost frontier of LLMs. Our models deliver state-of-the-art performance per FLOP, achieving competitive results at a fraction of the training cost. By optimizing architecture, data strategy, and training efficiency end-to-end, we prove that frontier-level capability does not require frontier-level spend.
We push the performance–cost frontier of LLMs. Our models deliver state-of-the-art performance per FLOP, achieving competitive results at a fraction of the training cost. By optimizing architecture, data strategy, and training efficiency end-to-end, we prove that frontier-level capability does not require frontier-level spend.
Bringing LLMs to Real World
Bringing LLMs to Real World
Bringing LLMs to Real World
We work directly with industry leaders to deploy our models in high-stakes workflows. We collaborate with partners across industries to develop domain-specific agents that support mission-critical decision-making—where accuracy, robustness, and accountability are non-negotiable.
We work directly with industry leaders to deploy our models in high-stakes workflows. We collaborate with partners across industries to develop domain-specific agents that support mission-critical decision-making—where accuracy, robustness, and accountability are non-negotiable.
We work directly with industry leaders to deploy our models in high-stakes workflows. We collaborate with partners across industries to develop domain-specific agents that support mission-critical decision-making—where accuracy, robustness, and accountability are non-negotiable.
Our models
Our models
Our models
All of our models are built entirely from scratch, by a team of just 8 researchers.
All of our models are built entirely from scratch, by a team of just 8 researchers.
All of our models are built entirely from scratch, by a team of just 8 researchers.
Tri-7B
A smaller, lighter-weight base model able to run on local compute and is optimized for Korean performance.
Tri-7B
A smaller, lighter-weight base model able to run on local compute and is optimized for Korean performance.
Tri-7B
A smaller, lighter-weight base model able to run on local compute and is optimized for Korean performance.
Tri-21B
Our flagship model achieving strong results in programming, reading comprehension, and math problem
Tri-21B
Our flagship model achieving strong results in programming, reading comprehension, and math problem
Tri-21B
Our flagship model achieving strong results in programming, reading comprehension, and math problem
Tri-70B
The largest dense model trained from scratch in Korea, with all intermediate checkpoints openly released.
Tri-70B
The largest dense model trained from scratch in Korea, with all intermediate checkpoints openly released.
Tri-70B
The largest dense model trained from scratch in Korea, with all intermediate checkpoints openly released.
Latest research
Latest research
Latest research
Our research team investigate new architectures of language models to increase utility and efficiency
Our research team investigate new architectures of language models to increase utility and efficiency
Our research team investigate new architectures of language models to increase utility and efficiency
Jan 1, 2026
Jan 1, 2026
Jan 1, 2026
Publication
Publication
Publication
Introducing rBridge
Introducing rBridge
Introducing rBridge
Introducing rBridge: The Most Efficient Path to Pre-training a Reasoning Model
Introducing rBridge: The Most Efficient Path to Pre-training a Reasoning Model
Introducing rBridge: The Most Efficient Path to Pre-training a Reasoning Model
September 10, 2025
September 10, 2025
September 10, 2025
Model Release
Model Release
Model Release
Tri Series Intermediate Checkpoints Release
Tri Series Intermediate Checkpoints Release
Tri Series Intermediate Checkpoints Release
Tri Series Intermediate Checkpoints Release
Tri Series Intermediate Checkpoints Release
Tri Series Intermediate Checkpoints Release
March 18, 2025
March 18, 2025
March 18, 2025
Model Release
Model Release
Model Release
Introducing Tri-7B
Introducing Tri-7B
Introducing Tri-7B
Introducing Tri-7B
Introducing Tri-7B
Introducing Tri-7B
Help us build the future
Help us build the future
Help us build the future
If you're interested in research at the frontier of what’s possible in AI, consider joining us.
If you're interested in research at the frontier of what’s possible in AI, consider joining us.
If you're interested in research at the frontier of what’s possible in AI, consider joining us.
Join our newsletter
Get the latest AI news and insights every week