This Startup Claims It Can Leapfrog OpenAI With a 10 Million-Token AI Model

Refiant AI, a California startup founded by South African mathematician Dr Viroshan Naicker, is claiming to have compressed a 120-billion-parameter model to run on a MacBook Pro. It says it has launched Protea, an LLM it claims has a 10 million token context window, nearly 10 times Claude and ChatGPT.

AI startup Refiant on Wednesday claimed to have launched a large language model with a multimillion-token context window, positioning it against much larger rivals OpenAI and Anthropic as competition intensifies over how much information AI models can process in a single prompt.

The California-based startup told Impact Newswire its Protea model can handle up to 10 million tokens, roughly equivalent to 7.5 million words or about 15,000 pages of text. The company said the model is available immediately without a waitlist.

The claim far exceeds the publicly available context windows offered by industry leaders. OpenAI’s latest GPT models and Anthropic’s Claude models support context windows of around 1 million tokens, while Google has also expanded its Gemini models to 1 million tokens as AI companies race to improve long-context reasoning.

A larger context window allows AI systems to analyze bigger datasets, such as entire codebases, legal archives or years of company documents, without splitting them into smaller chunks. However, bigger context windows do not necessarily translate into better performance, with many models still struggling to accurately retrieve information buried deep within lengthy prompts.

Refiant said Protea overcomes the so-called “lost in the middle” problem, a known limitation in long-context AI models where performance deteriorates on information located in the middle of very long inputs. Impact Newswire could not independently verify the company’s claims.

“Long-context AI has been talked about for over a year now, but hasn’t really been commercially available,” Chief Executive and co-founder Viroshan Naicker told this reporter.

Unlike OpenAI, Anthropic and Google, which have largely scaled AI performance using increasingly powerful computing infrastructure, Refiant says its approach focuses on computational efficiency. The startup adds that it uses optimization techniques inspired by evolutionary biology and swarm intelligence rather than relying primarily on additional hardware.

The company claims it previously compressed a 120-billion-parameter open-source model to run on an Apple MacBook Pro with 18 GB of memory, helping it secure a $5 million seed funding round led by VoLo Earth Ventures.

“Customers don’t need more waitlists,” said Mathew Haswell, Refiant’s chief product and operating officer in a press statement. “They need models they can test, break and build with. Protea is live, and we want people to use it from day one.”

Refiant adds it has also demonstrated an internal prototype with a 100 million-token context window, but failed to disclose when such a model would become commercially available.

The announcement comes as AI developers increasingly compete on context length alongside reasoning ability and multimodal capabilities, with enterprise customers seeking models capable of processing vast corporate datasets while maintaining accuracy. Independent evaluations have shown that simply increasing context size does not always improve performance, making reliable long-context reasoning one of the industry’s remaining technical challenges.

Stay ahead of the stories shaping our world. Subscribe to Impact Newswire for timely, curated insights on global tech, business, and innovation all in one place.

Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.

Got a story to share? Pitch it to us at info@impactnews-wire.com and reach the right audience worldwide