A token is roughly three quarters of an English word, so Refiant's 10M token window fits around 7.5M words in one pass — the 'lost in the middle' failure where long context models drift on middle sections is what Protea claims to fix.
Ask a chatbot to summarise a long document and watch what happens around the middle. If the document is a 200-page regulatory filing, a sprawling codebase, or a decade of customer emails, the answer usually starts and ends cleanly and then slips, quoting the wrong clause, missing one entirely, or quietly inventing a third. Researchers call the failure the "lost in the middle" problem: models stay accurate at the start and end of a long input, then drift in between.
Refiant, a small South African startup, says its new model Protea is built around that specific failure. The flagship tier accepts a 10-million-token context window and is available immediately, with no waitlist, the company writes.
A token is roughly three-quarters of an English word, so Refiant's 10-million-token window is about 7.5 million words in one prompt, roughly 15,000 pages back-to-back by techfinancials's count. Refiant's framing of the use case is 20 to 30 years of one person's documents, five years of email and Slack, decades of clinical-trial data, or a regulatory archive, ingested at full fidelity in a single pass.
That is the architectural bet. Anthropic, Google DeepMind, and the long-context specialist Magic have all been publishing or hinting at progressively larger windows over the last two years; the dominant pattern has been more compute, longer attention spans, bigger positional encodings, more careful training. Refiant's pitch is a different axis. The team describes its training as swarm-optimization, a biologically inspired, agent-based search across model configurations rather than a single brute-force run. The claim is that this approach finds configurations that hold accuracy across very long contexts at a smaller energy and hardware footprint than the brute-force pattern.
The energy angle is part of why VoLO Earth Ventures, a climate-focused fund, led Refiant's $5 million seed round. The startup is also partnering with Imperial College London and University College London's Sargent Centre for Process Systems Engineering on compression work for edge deployment, running serious models on laptops and small devices rather than in large data centres. The team previously compressed OpenAI's GPT-OSS-120B to run on a MacBook Pro with 18GB of RAM, a useful provenance point for anyone asking whether the swarm-optimization story holds up outside the company.
Refiant is a $5 million seed-stage company with three co-founders: Dr Viroshan Naicker, Siddharth Gutta, and Mathew Haswell. "First to ship at this scale" and "successfully tackles the lost in the middle problem" are the company's own claims, carried verbatim by IT-Online's coverage of the launch. techfinancials echoes those claims as "10x the memory of Claude and ChatGPT," which is Refiant's competitive positioning repeated by a sympathetic outlet, not an independent benchmark. No neutral third-party evaluation has been published for Protea: no Hugging Face leaderboard entry, no academic needle-in-a-haystack test, no MMLU-Pro long-context variant.
Whether the swarm-optimization approach is a genuine second axis for long-context training, or a one-off result that does not hold up outside the company's own evaluations, is a question only adversarial benchmarks can resolve. Magic and Anthropic have kept shipping longer-context claims on roughly the same timeline; Google's Gemini team has extended its window on a similar one. The next trigger is the first adversarial benchmark on Protea: an independent needle-in-a-haystack result, an open third-party leaderboard entry, or a published replication of the swarm-optimization method.