AI optimization startup Refiant launched Protea, a suite of long‑context models with a 10 million‑token window, the largest publicly available yet.
AI optimization startup Refiant has unveiled Protea, a new suite of long‑context language models that can process up to 10 million tokens in a single prompt, making it the largest publicly available context window to date.
Why a 10‑million‑token window matters
Long‑context capabilities enable developers to feed entire documents, codebases, or multimodal datasets into a model without chopping them into smaller fragments, reducing the need for complex prompt engineering and preserving context continuity.
Compared with rival offerings that typically cap at 32 k or 100 k tokens, Protea’s window opens up use cases such as full‑book summarization, exhaustive legal review, and comprehensive scientific literature analysis.
Technical highlights of Protea
- 10 million‑token context window, configurable per model tier
- Optimized transformer architecture to maintain latency under 5 seconds for 1‑million‑token inputs
- Hybrid retrieval‑augmented generation that indexes external data on‑the‑fly
- Open‑source inference code released under the Apache 2.0 license
Refiant claims the model scales linearly in memory usage thanks to a sparse attention mechanism, allowing it to run on clusters of commodity GPUs rather than requiring specialized hardware.
Market positioning and competition
While giants like OpenAI and Anthropic have hinted at multi‑million token windows in internal roadmaps, none have yet released a publicly accessible model of this size. Refiant’s move positions it as a first‑mover in the emerging “ultra‑long‑context” niche.
The company plans to monetize Protea through a tiered API, offering free access for research and paid plans for enterprise workloads that demand the full 10‑million token capacity.
We wanted to break the token ceiling that has limited real‑world applications of LLMs for years, Refiant CEO Maya Patel said in the launch announcement.
Analysts see the development as a potential catalyst for new AI products that can handle entire knowledge bases without external chunking, though they caution that cost and latency will remain key challenges for widespread adoption.