en.Wedoany.com Reported - South African AI optimization company Refiant has released a long-context AI model named Protea, featuring a context window of up to 10 million tokens—ten times larger than existing AI tools like ChatGPT and Claude, making it one of the largest publicly available context windows to date.
The context window of a large language model determines the amount of text (measured in tokens) the model can "remember" at any given time. A larger context window enables AI models to process longer inputs and incorporate more information into each output. Currently, mainstream LLMs such as Claude, Google Gemini, and OpenAI's GPT series have context windows ranging from thousands to 1 million tokens. Protea's 10 million token window can accommodate the equivalent of 7.5 million words or 15,000 pages of content, allowing it to comprehend up to five years of emails or 83 novels in a single pass.
Refiant noted that previous state-of-the-art models struggled when processing working memory exceeding several hundred thousand tokens, forcing users to adopt complex workarounds to compensate for content the model could not access. Protea can process entire regulatory archives, corporate codebases, and decades of clinical trial data in one go without the need for segmentation. The company stated that this model is the first of its kind to be deployed at such a scale in production, delivering capabilities that competitors have only promised so far. Protea is also available to individual users, but is primarily being launched as a way for enterprises to handle truly large-scale, complex information. Users can start developing on versions ranging from 1 million to 10 million tokens immediately, with no queues or approval processes, and it is completely free.
In enterprise applications, engineering teams can use Protea to digest an entire codebase in one go, compressing a month's worth of analysis into a single day. Insurance companies can upload years of claims data to detect fraud patterns, and law firms can input five years of case files to identify relevant precedents. For teams building agent workflows, the model enables agents to operate within a vast context without losing early reasoning threads. Refiant explained that Protea solves the "lost in the middle" problem, a recognized limitation where models with context windows exceeding one million tokens lose information in the middle portion.
Refiant was founded in 2025 by Dr. Viroshan Naicker, Siddharth Gutta, and Mathew Haswell. Its approach draws on evolutionary search and swarm optimization, mimicking natural systems to find efficient solutions. The company first applied these techniques to model compression, shrinking OpenAI's GPT-OSS-120B model to run on a MacBook Pro with 18GB of RAM. In April 2026, Refiant secured a $5 million seed round led by VoLo Earth Ventures and established research collaborations with Imperial College London and UCL's Sargent Centre for Process Systems Engineering.
Refiant is registered in Delaware, USA, and the Protea model is named after South Africa's national flower. The founders emphasize building AI using nature-inspired methods, prioritizing efficiency over brute-force computation. Naicker, co-founder and CEO, stated that long-context AI has been discussed for over a year but has not yet been truly commercialized, and customers need models they can test and build on. Haswell, Chief Product Officer, added that Protea is live and they want people to start using it from day one.
This beta release is the first phase of a three-stage roadmap, with more announcements expected over the next three months. The platform is currently free, with plans for a commercial version in the long term. Protea offers versions with 1 million, 5 million, and 10 million tokens, with access requiring no queues or approval processes. The company has internally demonstrated a working prototype with a 100 million token context window and is exploring how to bring it to production at that scale.
