Voltropy

Artificial Intelligence · Large Language Models

Launch

Voltropy launches Vast-10M, first LLM with 10-million-token context

A new attention mechanism lets the model get smarter at long context instead of degrading, unlike prior scaling tricks.

• Voltropy unveiled Vast-10M, a frontier AI model family with a native ten-million-token context window, available today via early access.
• Voltropy is an AI research lab based in Orange County, California, co-founded by CEO Clint Ehrlich and CTO Ted Blackman, previously known for its Lossless Context Management agent architecture.
• Vast-10M comes in three sizes built on existing open models: Flash (DeepSeek V4 Flash), Medium (GLM 5.2), and Pro (DeepSeek V4 Pro), all extended via a new algorithm called Voltropy Scalable Attention (VSA).
• At the one-million-token BEAM benchmark tier, Vast-10M-Flash outscores Anthropic's Fable 5.1 by 10.68% and matches OpenAI's GPT-6 Astra, then keeps scaling to 10 million tokens where rivals stop.
• Vast-10M-Flash retains 82.49% of its one-million-token BEAM score at ten million tokens, and even scores higher at 10M tokens (40.20) than its base DeepSeek model does at 1M tokens (39.69).
• Frontier context windows had been stuck near one million tokens for roughly three years, with OpenAI's GPT-6 Astra at 1.05 million and Anthropic's Fable capped similarly, while other labs like MiniMax and NaiveAI just shipped their own million-token models this same week.
• Prior methods for extending context windows typically traded away model intelligence, so a technique that boosts both scale and reasoning quality simultaneously could reset expectations for how much raw data enterprise AI tools can ingest at once.