A Cape Town startup just gave Claude and Gemini a 10-million-token problem to solve

Refiant AI's new Protea model can hold roughly 7.5 million words in a single prompt, free and without a waitlist — a long-context bet built on compression, not bigger data centres.

Ten million tokens. That is roughly 7.5 million words, or about seventy copies of a typical 300-page novel, sitting in a single AI prompt at once. It is also the number a small Cape Town-founded startup, Refiant AI, put on the table this week — and it is roughly twenty times the working memory Anthropic offers on Claude's enterprise tier, and ten times what Google gives most Gemini users.\n\nThe model is called Protea, named for South Africa's national flower, and it launched on July 8 in three versions — with context windows of one million, five million, and ten million tokens. Anyone can use it, for free, on Refiant's platform, with no waitlist and no approval process. "Customers don't need more waitlists," cofounder Mathew Haswell said in the company's announcement. "They need models they can test, break, and build with."\n\nWhat a context window actually buys you\n\nA model's context window is its working memory — how much text it can hold in mind at once before it has to forget the beginning to make room for the end. A short window forces a workaround familiar to anyone who has used AI on a long document: split the file into chunks, feed them in piece by piece, and hope the model doesn't lose the thread between chunks. A ten-million-token window removes that problem for all but the largest jobs. Refiant says a law firm could run due diligence across hundreds of contracts in one pass; an insurer could analyse years of claims history without re-querying the model each time; an engineering team could load an entire software codebase and ask questions about all of it at once.\n\nThis is not Refiant's first act. The company — founded in 2025 by quantitative mathematician Viroshan Naicker, finance-trained Siddharth Gutta, and commercial lead Mathew Haswell — raised a $5 million seed round in April, led by California's VoLo Earth Ventures, on the strength of a different but related trick: compressing a 120-billion-parameter model down to run on a standard 12-gigabyte laptop, a task that would normally demand at least 80 gigabytes of memory, while keeping 95 to 99 percent of the original performance. The pitch to VoLo, a climate-focused fund, was about energy: Microsoft and Meta alone committed close to $50 billion each to new data-centre leases in the first quarter of 2026, and Refiant's founders argue that making models smaller and more efficient is a more sustainable answer than building more power-hungry infrastructure to run bigger ones.\n\nProtea is that same compression logic pointed at memory instead of parameter count. Naicker, speaking to CNBC Africa, described the underlying insight plainly: "Having got results that say compression is interesting and it works — well, if you can compress some things, then you can also expand other things." The company says it already has a working internal prototype handling 100 million tokens and is now figuring out how to benchmark and productionise it, with Protea positioned as only the first of three planned releases.\n\nA claim worth testing, not just repeating\n\nRefiant has not published independent benchmark results for Protea. Naicker told reporters the company has internal numbers on recognised long-context tests but chose not to release them, framing the free, waitlist-free launch itself as the proof: try it, break it, see what it does. That is a reasonable posture for a startup trying to win developer trust quickly, but it means the ten-million-token figure and Refiant's latency claims are, for now, self-reported. Enterprises weighing Protea for anything sensitive — clinical data, financial records, legal discovery — would be right to run their own tests alongside the company's rather than take the number on faith. It is also worth noting Refiant is not alone at this frontier: a rival called Subquadratic reportedly reached a twelve-million-token window back in May, two months before Protea shipped, making this less a solo breakthrough than a real, fast-moving contest among specialised firms racing past the token counts offered by the household names.\n\nWhy it lands differently in Africa\n\nThe part of this story that matters most for African builders is not the number itself — it's where the number came from. Long-context research has largely been the preserve of labs with the capital to run enormous training clusters. Refiant's approach, by its founders' own account, gets there through algorithmic efficiency rather than brute-force scale, developed by a team split between South Africa and outposts in the US, Cambridge, and Sardinia. For African fintechs, insurers, and public agencies that have neither the budget for premium enterprise API tiers nor reliable access to the compute those tiers assume, a free, high-capacity model built by people who understand African infrastructure constraints first-hand is a meaningfully different starting point than yet another discount tier bolted onto a Silicon Valley product. Whether Protea's numbers survive independent scrutiny is still an open question. That the challenge to Claude and Gemini's memory ceiling came from Cape Town, not San Francisco, is not.