Refiant, an Artificial intelligence optimisation startup has launched Protea, a new family of long-context AI models featuring a context window of up to 10 million tokens.
This positions the startup among a small group of companies pushing the boundaries of how much information AI systems can process at once.
Viroshan Naicker, chief executive and co-founder of Refiant, said long-context AI has been discussed extensively over the past year but has remained largely inaccessible for commercial use.
“Long-context AI has been talked about for over a year now, but hasn’t really been commercially available,” Naicker said.
The company said in a note that the models are immediately available through its platform without a waitlist or approval process, allowing developers and enterprises to access one million-, five million-, and 10-million-token versions for free.
The launch comes as AI companies race to extend context windows with the amount of text, code, or other data an AI model can retain and reason over in a single interaction.
While most commercially available large language models support context windows ranging from hundreds of thousands to around one million tokens, larger windows promise to eliminate the need to split massive datasets into smaller chunks before analysis.
According to Refiant, its 10-million-token model can process approximately 7.5 million words, equivalent to around 15,000 pages of text in a single prompt. The company said this would enable enterprises to analyse years of documents, software codebases, clinical research records, legal archives and internal communications without breaking them into smaller datasets.
The startup said the expanded context window could significantly reduce the time required for complex analysis. Engineering teams, for example, could review entire software repositories in one pass, while insurers could analyse years of claims records without repeatedly querying the model.
Refiant also claims Protea addresses the so-called ‘lost in the middle ’ problem, which is a known limitation in long-context AI systems where models struggle to accurately retrieve or reason over information located in the middle of extremely large prompts, despite performing well on information at the beginning or end.
Mathew Haswell, co-founder and chief product officer at Refiant, said customers need AI models they can immediately experiment with rather than waitlists for early access.
The company was founded by Naicker alongside Siddharth Gutta and Haswell.
Before launching Protea, the startup focused on AI model optimisation, claiming it successfully compressed GPT-OSS-120B to run on a laptop equipped with 18GB of RAM.
The work helped the company secure a $5 million seed funding round led by VoLo Earth Ventures and establish research partnerships with Imperial College London and the UCL Sargent Centre for Process Systems Engineering.
Looking ahead, Refiant said it has already demonstrated an internal prototype capable of handling a 100-million-token context window and is exploring how to benchmark and commercialise the technology.
The launch shows intensifying competition in the AI industry, where model developers are increasingly focusing on expanding context windows alongside improvements in reasoning, multimodal capabilities and efficiency.
Get Newsletter Updates
Enjoying our column?
Subscribe to our specialised **Tech Pulse** feed to receive fresh reports and analyses directly in your inbox.
