FrontiersMind

Open research toward Sovereign Superintelligence.

We are a research lab working toward sovereign superintelligence. We believe the most capable models should be efficient enough for every organisation to own and run on its own hardware. So we design new architectures, pre-train our own models from scratch, and release our papers, weights and code openly.

Papers

• Grouped Value Attention · Sep 2026: about 45% fewer KV cache, with near GQA accuracy (FlashGVA coming soon).
• Grouped Query Experts · Jun 2026: matches all-active GQA with half the query heads active.

Website · Blog · Discord