New Frontier in Long-Context AI: Protea Boasts 10 Million Tokens of Working Memory

Refiant, a South Africa-based startup specializing in AI optimization, has unveiled its Protea family of models with memory capacities that significantly surpass industry standards. The top-tier version offers an impressive 7.5 million words—approximately 15,000 pages—in a single pass, dwarfing competitors like Anthropic’s Claude (with 500,000 tokens) and Gemini (with 1 million tokens).

Addressing the Attention Span Challenge

The Protea models tackle one of AI’s most persistent limitations: context window size. Even advanced frontier models typically lose accuracy after processing a few hundred thousand tokens, requiring developers to break down complex inputs into smaller chunks. With Protea, Refiant claims users can process entire regulatory filings, codebases, or years of correspondence without fragmentation.

This expanded memory allows for new applications—law firms could run due diligence across hundreds of contracts in one pass, while insurers could analyze complete claims histories without repeatedly querying the model.

A Competitive Landscape

While Refiant’s announcement positions Protea as a leader in long-context AI, it’s part of a growing race among specialized firms. Subquadratic previously launched a 12-million-token window earlier this year, demonstrating increasing competition in this space.

Notably, Refiant hasn’t yet released independent benchmark results, instead inviting developers to test the models directly—a common approach for startups seeking real-world adoption rather than relying solely on published metrics.

Nature-Inspired Computing Approach

Refiant’s technology draws inspiration from biological systems. Their founders apply evolutionary search and swarm optimization techniques—similar to how ants or bees solve complex problems collectively—to create efficient AI models.

The company previously gained attention for compressing OpenAI’s 120-billion parameter GPT model to run on a MacBook Pro, requiring only 12-18GB of RAM. This achievement secured them a $5 million seed round and research partnerships with leading universities.

Accessible Without Gatekeepers

In contrast to the waitlist culture surrounding many recent AI product launches, Refiant is offering all three tiers of Protea—ranging from 1 million to 10 million tokens—completely free through their website without requiring approval.