South Africa’s Refiant claims its AI model, Protea, has 20 times Claude’s memory

Refiant, a South African-founded startup, has just put a genuinely striking number on the table: it offers 10 million tokens of working memory, roughly 20 times what Anthropic’s Claude on its enterprise tier.

It’s a real technical feat if it holds up. And, it comes free, live, and without a waitlist.

The AI optimisation company behind an earlier and widely discussed trick that squeezed a 120-billion-parameter OpenAI model onto a MacBook Pro launched its Protea family of models this week. The top-tier version holds around 7.5 million words, or roughly 15,000 pages, in a single pass.

That’s a meaningful jump over the 500,000-token ceiling on Claude’s enterprise offering and well past Gemini’s one-million-token top tier.

This is more important considering that one of the biggest challenges of AI has been attention span. Even the strongest frontier models typically start losing accuracy after a few hundred thousand tokens, forcing developers to chop regulatory filings, codebases, or years of correspondence into fragments.

Refiant says Protea does away with that entirely, letting a law firm run due diligence across hundreds of contracts in one pass or an insurer work through years of claims history without re-querying the model each time.

It’s a strong claim, and it’s worth pairing with some context before enterprises start building on it.

Refiant isn’t quite first to this number

Refiant’s announcement describes its 10-million-token window as “one of the largest ever made publicly available”, which is broadly fair but slightly understates the field. Tech outlet, The New Stack, noted that a rival called Subquadratic had already reached further: debuted a 12-million-token window back in May, two months ahead of Protea’s launch.

That’s less a knock on Refiant’s achievement than a useful framing.

Long-context AI has quietly become a genuine race among smaller, specialised firms, and Refiant has produced one of the strongest entries in that field rather than the only one.

The one thing worth flagging clearly is that Refiant hasn’t published independent benchmark results yet.

Co-founder and CEO Viroshan Naicker said the company has internal results on recognised long-context tests but chose not to release them, framing it instead as an invitation for developers to try the models themselves rather than take Refiant’s word for it.

That’s a fair approach for a startup building hands-on adoption, and it’s consistent with how Refiant has generally talked about its work.

It means, though, that the 10-million-token figure and the latency claims are currently all self-reported, with no third-party replication on record yet. For enterprises evaluating Protea on anything sensitive, from clinical trial data to financial records, that’s a reasonable prompt to run their own tests alongside the company’s, rather than a reason for scepticism about the underlying work.

More about Refiant

Refiant’s approach leans on what its founders call nature-inspired computing. Evolutionary search and swarm-style optimisation are modelled on how ants, bees, and other collective systems solve complex problems efficiently without centralised control.

Naicker, a quantum mathematician by training, co-founded the company in 2025 alongside Siddharth Gutta, who has a finance background, and Mathew Haswell, who leads commercial scaling as chief product and operating officer.

The company first drew attention for compressing OpenAI’s GPT-OSS-120B model down to run on a MacBook Pro, with reported figures putting the RAM requirement somewhere between 12GB and 18GB. That result was enough to pull in a $5 million seed round led by VoLo Earth Ventures, plus research partnerships with Imperial College London and UCL’s Sargent Centre for Process Systems Engineering.

Protea is the company’s first major commercial product since that raise, and it comes in three tiers: one million, five million, and 10 million tokens, all accessible for free through refiant.ai without approval gates. 

Naicker argues the industry has talked about long-context AI for over a year without actually shipping it commercially at scale, and Haswell has framed the free, ungated access as a deliberate rejection of the waitlist culture that’s defined recent AI product launches.

Where this sit against the frontier labs

For context, Anthropic’s Claude currently offers context windows up to 500,000 tokens on certain enterprise tiers, while Google’s Gemini tops out around one million tokens on its higher-tier plans. If those figures hold, Protea’s top tier would represent a meaningful jump over what the established labs offer commercially today, even accounting for Subquadratic’s head start.

Whether a bigger context window translates into better real-world performance is a separate question entirely. The lost-in-the-middle problem, a situation where models handle information at the start and end of a long prompt well but lose track of what’s buried in between, has dogged every long-context model to date, and Refiant’s claim to have solved it is indeed a breakthrough.

Naicker has acknowledged latency as a genuine constraint too, saying that inference across a 10-million-token window inevitably carries a performance cost, though he maintains internal testing shows it’s manageable.

Refiant says it already has a working internal prototype handling 100 million tokens and is figuring out how to benchmark and productionise it. Wednesday's launch is described as the first of three planned phases, with more expected within three months.

For African enterprises and developers, the appeal is obvious: a free, immediately accessible model from a South African-founded company, built by a team with genuine technical pedigree, offering context capacity that could handle the kind of unwieldy datasets, years of regulatory filings, sprawling codebases, and multi-year claims histories that smaller teams often can’t afford to process through paid, gated frontier models.

That’s a real opportunity, particularly for fintech, insurance, and legal-tech operators across the continent looking to cut costs on data processing.