Artificial intelligence optimization startup Refiant Inc. has 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. The models went live on July 8, 2026, at refiant.ai with no waitlist and no approval process.
Context windows determine how much information a model can hold in working memory at once. Most leading AI models today 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 significantly. For ongoing coverage of AI model launches and industry trends, readers can find the latest AI news at our homepage.
What 10 Million Tokens Means in Practice
A 10 million-token context window works out to roughly 7.5 million words, or about 15,000 pages of text. Refiant frames the practical implications in stark terms: an entire enterprise codebase can be loaded at once, as can a full regulatory archive or years of clinical trial data that previously had to be chopped up and fed to a model piece by piece.
The company puts it even more concretely. Ten million tokens can hold 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 analysis in a day. An insurer could sift through years of claims without first tearing the data into chunks.
Protea is available in three versions spanning 1 million, 5 million, and 10 million tokens, giving developers flexibility to match the context size to their workload and budget.
Tackling the Lost-in-the-Middle Problem
One of the most persistent challenges with large context windows is the so-called lost-in-the-middle problem, a documented weakness where models stay accurate at the beginning and end of an input but lose track of material buried between those endpoints. Refiant says Protea specifically addresses this issue, though the company has not yet published detailed benchmark results demonstrating the improvement.
The lost-in-the-middle problem has been a significant obstacle to the practical adoption of long-context models. Even when a model technically accepts millions of tokens, its reliability on information in the middle of that context has often been questionable, limiting real-world utility for tasks like full-codebase analysis.
Founded on Swarm Optimization and Evolutionary Search
Refiant was founded in 2025 by Viroshan Naicker, Siddharth Gutta, and Mathew Haswell. Naicker is a quantum mathematician, while his co-founders come from finance and commercial scaling backgrounds. The team argues 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, the same collective intelligence principles observed in flocks of birds, schools of fish, and ant colonies. The company first applied this approach to model compression, successfully shrinking OpenAI's GPT-OSS-120B to run on a MacBook Pro with just 12 gigabytes of RAM. Those compression results helped secure research partnerships with Imperial College London and University College London's Sargent Centre for Process Systems Engineering.
Funding and Future Plans
The Protea launch follows a $5 million seed round that Refiant closed early in 2026, led by VoLo Earth Ventures. The South African-founded company is now based commercially in the United Kingdom.
Chief Executive Viroshan Naicker noted that long-context AI has been talked about for over a year but has not really been commercially available. Co-founder and Chief Product Officer Mathew Haswell added that customers want models they can test and build with rather than more waitlists, a direct criticism of the gated access approach adopted by many larger AI labs.
Refiant said it has already run an internal prototype at 100 million tokens, though how to benchmark that and turn it into a product remains an open question. The Protea release is the first of three planned stages, with the company promising more within three months.
The Long-Context AI Race Heats Up
Refiant enters an increasingly competitive long-context AI market. While major labs have gradually expanded context windows, most commercially available models still operate well below the 10 million-token threshold. Refiant's decision to launch with no waitlist and immediate access represents a contrast to the staged rollout approach common among larger competitors.
The company's swarm optimization methodology also distinguishes it from rivals who rely primarily on conventional transformer scaling. Whether nature-inspired algorithms can deliver sustained performance advantages at production scale remains to be seen, but Refiant's compression track record and university partnerships lend credibility to the approach.
Stay Ahead of AI
Protea models are available now at refiant.ai. For more AI industry coverage and analysis, visit AI Buzz Wire.
Read more AI news →
