Key Takeaways
- Context window: the Protea model reaches 10 million tokens, equivalent to roughly 15,000 pages of text processed in a single pass.
- Technology: an architecture built on evolutionary and swarm algorithms, developed by Refiant AI to cut energy consumption compared with standard Transformer models.
- Roadmap: an internal prototype already handles 100 million tokens, ten times beyond the publicly released version.
The Launch and Technical Specs
Refiant AI, a startup based in Pretoria, released the Protea suite on July 8, 2026. The system ships in three configurations: 1 million, 5 million, and 10 million tokens of context. The maximum tier absorbs the equivalent of 7.5 million words, enough to process up to five years of a single user's email correspondence. Access runs through the refiant.ai platform, with no waitlist.
Solving the "Lost in the Middle" Problem
Protea holds retrieval accuracy steady even on extremely long inputs, addressing the performance drop that plagues rival models when the requested information sits in the middle section of a document.
Corporate Structure and Funding
The company, founded in 2025 by Viroshan Naicker, Siddharth Gutta, and Mathew Haswell, closed a seed round led by VoLo Earth Ventures three months before the public launch. The internal 100-million-token prototype signals the next phase of the technical roadmap.
