The breakthrough relies on Voltropy Scalable Attention, a proprietary algorithm designed to expand context windows without the intelligence degradation typical of previous scaling methods. According to CTO Ted Blackman, the architecture actually enhances performance at shorter lengths, allowing the model to compete with top-tier benchmarks while maintaining full recall capabilities across its massive ten-million-token range. The company is rolling out the technology in three distinct versions: Vast-10M-Flash, Medium, and Pro, which build upon existing DeepSeek and GLM foundations.

Performance metrics on the industry-standard BEAM benchmark indicate the Flash variant achieves parity with OpenAI’s GPT-6 Astra while surpassing Anthropic’s Fable 5.1 at the one-million-token threshold. By maintaining high recall at maximum capacity, the model aims to move beyond the limitations that have constrained frontier AI development for the past three years. Early access to the model family is currently available through the company’s website.