A South African-founded startup has thrown down a marker in the global race to build AI models with longer memories. Refiant AI this week launched Protea, a suite of large language models that the company says can process up to 10 million tokens — the chunks of text an AI reads and writes — in a single prompt, one of the largest context windows made publicly available so far.
A bigger context window lets a model take in more information before it answers, rather than forgetting or losing track partway through. Protea comes in three versions — one million, five million and 10 million tokens — and is free to use with no waitlist or approval. At full stretch, Refiant says the 10-million-token model can hold roughly 7.5 million words at once, enough to read hundreds of contracts, years of claims data or an entire software codebase in a single pass.
10 millionTokens Protea can process in one prompt~7.5 millionWords it can hold at maximum capacity
For scale, Refiant says most mainstream tools trail well behind: Anthropic's Claude supports up to 500,000 tokens on certain enterprise plans, while Google's Gemini offers up to one million on its higher tiers.
Built by a local team
Founded in 2025 by Mathew Haswell, Viroshan Naicker and Siddharth Gutta — a team spanning mathematics, finance and commercial scaling — Refiant builds systems to compress AI models and cut the cost of running them. The launch follows a $5 million seed round led by VoLo Earth Ventures in April, and is the firm's first major product since.
"Long-context AI has been talked about for over a year now, but hasn't really been commercially available," said Naicker, the company's chief executive. Co-founder Haswell added: "Customers don't need more waitlists. They need models they can test, break and build with." The company says it has already demonstrated an internal prototype capable of 100 million tokens, and that Protea is the first of three planned releases.
What it means
For South Africa's tech scene, Protea is a rare example of a locally founded company competing at the frontier of AI rather than simply adopting tools built elsewhere. If it performs as advertised, it could be genuinely useful to the country's law firms, insurers and engineering teams that need to reason over huge documents without chopping them into pieces. The claims are bold, though, and the real test will be independent benchmarks — long-context models have a habit of losing accuracy on detail buried deep in the text.
