Artificial intelligence optimization startup Refiant Inc. today launched Protea, a suite of long-context AI models led by a 10 million-token context window that the company says ranks among the largest ever made publicly available.

Context windows determine how much information a model can hold in working memory at once. Most leading models handle no more than a few hundred thousand tokens before performance degrades, forcing developers into workarounds that break large datasets into fragments.

Refiant says Protea lifts that ceiling. An entire enterprise codebase can go in at once. So can a full regulatory archive, or years of clinical trial data that used to be chopped up and fed to a model piece by piece.

A 10 million-token window works out to roughly 7.5 million words, or 15,000 pages. Refiant puts it in blunter terms: up to five years of one person’s email and Slack messages, or decades of their reports and files. An engineering team could load a full codebase and run its analysis in a day. An insurer could sift years of claims without first tearing the data into chunks.

The company said Protea also tackles the “lost in the middle” problem, a documented weakness of large context windows in which models stay accurate at the beginning and end of an input but lose track of material buried between.

Refiant was founded in 2025 by Viroshan Naicker, Siddharth Gutta and Mathew Haswell. Naicker is a quantum mathematician; the others come from finance and commercial scaling. They argue that modern AI models are deeply inefficient and that better methods already exist in nature. Refiant’s approach borrows from evolutionary search and swarm behavior.

Refiant first applied the approach to model compression, shrinking OpenAI Group PBC’s GPT-OSS-120B to run on a MacBook Pro with 12 gigabytes of RAM. Those results helped secure research partnerships with Imperial College London and University College London’s Sargent Centre for Process Systems Engineering.

“Long-context AI has been talked about for over a year now but hasn’t really been commercially available,” Chief Executive Viroshan Naicker said. Co-founder and Chief Product Officer Mathew Haswell added that customers want models they can test and build with rather than more waitlists.

Refiant said it has already run an internal prototype at 100 million tokens. How to benchmark that and turn it into a product is still an open question. The launch is the first of three planned stages. It said more is coming within three months.

The Protea release follows a $5 million seed round Refiant closed early this year led by VoLo Earth Ventures.

The models are live now at refiant.ai, with versions spanning 1 million, 5 million and 10 million tokens and no waitlist or approval process to access them.

Image: Refiant