Protea: A New Contender in Long-Context AI
Refiant, a South African startup specializing in AI optimization, has launched its Protea family of models boasting impressive memory capabilities. The top-tier version offers 10 million tokens of working memory—approximately 20 times what Anthropic’s Claude provides on its enterprise tier.
Key Features and Benefits
The Protea model can process around 7.5 million words or 15,000 pages in a single pass. This expanded context window addresses a common challenge with AI models that typically lose accuracy after processing only a few hundred thousand tokens. With Protea:
- Law firms can analyze hundreds of contracts simultaneously for due diligence
- Insurers can process years of claims data without re-querying the model
- Developers can work with entire codebases in one prompt
Competitive Landscape
While Refiant’s announcement positions Protea as having one of the largest publicly available context windows, it’s worth noting that Subquadratic already debuted a 12-million-token window earlier this year. This highlights a growing race among specialized firms to push the boundaries of long-context AI.
Refiant hasn’t yet released independent benchmark results, but co-founder Viroshan Naicker encourages developers to test the models themselves rather than relying solely on company claims—a common approach for startups seeking hands-on adoption.
Refiant’s Approach
The company draws inspiration from natural systems like ant colonies and bee swarms to optimize AI performance without centralized control. This “nature-inspired computing” philosophy has already enabled them to compress large language models to run on consumer hardware, earning a $5 million seed round.
Protea is currently available in three tiers (1 million, 5 million, and 10 million tokens) through refiant.ai—all accessible for free without approval gates.