Trillion Labs launches with a plan to publish the AI research rivals keep private

A new AI lab called Trillium Labs has launched with an unusual pitch: it plans to show its work. Many frontier labs keep their riskiest research private.

Trillium says it intends to publish findings on two of the field’s most sensitive topics, AI self-improvement and model behavior.

What Trillium is promising

Self-improvement means research into AI systems that can make themselves better. Model behavior covers how AI systems actually act once they are trained, including whether they do what their builders intended, and what happens when they don’t.

Openness has company, including a similarly named startup

Trillium isn’t the only outfit pushing in this direction. South Korea’s Trillion Labs has built its brand around what it calls “radical openness.”

Trillion Labs was founded in 2024 by CEO Shin Jae-min, also known as Jay Shin, a former research scientist at Naver. Within one year, the company developed and open-sourced a family of language models it calls the Tri series.

Those models range from 0.5B to 70B parameters and were released under the Apache 2.0 license, which lets others use, modify and build on the code with few strings attached.

Trillion Labs released intermediate training checkpoints with the 70B model, which it describes as the first public release of its kind at that scale globally. Checkpoints are snapshots of a model taken partway through training. Sharing the in-between versions matters for studying how models learn and how their behavior emerges.

Trillion Labs has also focused on cost. Its Tri-21B model, open-sourced in July 2025, reportedly cut training costs to 1/12 of previous methodologies, with techniques such as XLDA cited as part of that efficiency push.

The company presented Korean language model work at NVIDIA GTC 2025, and in June 2026 announced a partnership with NVIDIA to develop industrial “world models” aimed at AI data centers and power infrastructure.

Trillion Labs has secured approximately $4-6 million in seed funding from investors including Kakao Ventures and Strong Ventures.

Why the openness debate matters now

Research on self-improvement and model behavior are precisely the questions regulators, safety researchers and the public most want answered, and also the questions companies are least eager to discuss in detail.

What this means

For smaller research teams, access to published research or to model checkpoints like those Trillion Labs released lets them study questions they otherwise could not touch. Cost-cutting methods of the kind Trillion Labs has pointed to may also make it more feasible for non-hyperscale players to compete in serious AI research.

Trillion Labs has paired its open releases with industry partnerships like its NVIDIA work. Whether Trillium finds a similar balance will determine if its openness is a lasting strategy or a launch-week talking point.