Our mission is to redesign how AI thinks, learns, and discovers.
We build architectures that bridge human-like reasoning and real-world impact—from specialized generalist models (Nirvana) and hardware-aware systems for medicine, to non-autoregressive diffusion models (SDAR) that mirror the iterative nature of the human mind. Building on these foundations, we created MARTI—a multi-agent system that is already driving breakthroughs in computational chemistry and enabling discoveries at a level worthy of Nature.
- Awesome-RL-LLMs – A curated list of RL for LLMs papers.
- Awesome-Diffusion-LLMs – A curated list of diffusion LLM papers.
- Awesome-Agentic-LongSequence-LLMs – A curated list of papers on long-sequence modeling for Agentic AI.
- PI-Mem - An Agentic Framework for Processing Ultra-Long Sequences.
- Nirvana – A specialized generalist model with task-aware memory mechanism.
- SDAR – A family of large diffusion language models (1.7B to 30B) combining diffusion and autoregression.
- SDAR-VL – The first large-scale block-wise discrete diffusion model for vision-language understanding.
- MARTI – A Framework for LLM-based Multi-Agent Reinforced Training and Inference.
- MacBench – A macOS benchmark for evaluating AI agents on real desktop workflows, featuring reproducible tasks, rule-based evaluation, and native app support for everyday scenarios.
- Arche-Harness- A novel multi-agent system enables highly sensitive computational chemistry discovery, delivering Nature-level performance.