Protea: A South African-Founded Startup Ships a 10 Million-Token Context Window

Refiant’s Protea models hold up to 10 million tokens in a single pass, about 7.5 million words or five years of messages, and the API is live with no waitlist.
artificial-intelligence
entrepreneurship
Author

Kabui, Charles

Published

2026-08-26

Keywords

long-context, context-window, 10-million-tokens, protea, south-africa

Download as PDF

Refiant, a startup founded in South Africa, launched Protea, a family of long-context models that hold up to 10 million tokens in a single pass. That is roughly 7.5 million words, about 15,000 pages, or five years of email and Slack messages. Protea is available now through an API or chat interface in 1 million, 5 million, and 10 million token windows, with no waitlist. Refiant was founded in 2025 by a quantum mathematician and two co-founders, and it builds on a compression technique that already shrank a 120-billion-parameter model to 12 GB of RAM while keeping 95-99% of its accuracy.

Whole codebases, regulatory archives, or years of claims data can now be fed to a model in one session instead of being chopped into chunks for a retrieval system. The South African founding is part of the story too: a model family at this scale shipping from a team based in Durban and Cape Town, with research partners at Imperial College London and UCL. Refiant’s own tests claim the compressed models are up to 100 times more energy efficient than data center hardware, though these are self-reported numbers and long-context models still lose accuracy on material buried in the middle of a window.

Context size is turning into a shipping race, not just a research one. Magic announced a 100-million-token model in 2024 and nothing public shipped, a pattern we covered with SubQ. Refiant put 10 million tokens in a free API within a year of founding. Efficiency, not just scale, is how a small team competes with labs spending billions.

Read More: AI memory: why bigger context windows don’t stop agents from forgetting

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Protea: {A} {South} {African-Founded} {Startup} {Ships} a 10
    {Million-Token} {Context} {Window}},
  date = {2026-08-26},
  url = {https://toknow.ai/posts/protea-refiant-10-million-token-context-window/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Protea: A South African-Founded Startup Ships a 10 Million-Token Context Window.” https://toknow.ai/posts/protea-refiant-10-million-token-context-window/.