Research blog
Posts from the Seele AI research team on the foundation models, systems, and benchmarks behind SeeleAgent — native 3D understanding and generation, multimodal architectures, real-time interactive world models, and how we evaluate game-development agents. New posts land here as our research progresses.
Games Must Be Played: Oasis and the Next Benchmark for Game Development Agents
Oasis is our next benchmark for game-development agents — capability-tree-first, Playground-based, and runtime-observed: it plays the result inside an engine instead of scoring static output.
READ POST →EVA01: Giving MLLMs a Native 3D Hand
A unified 3D-native multimodal model — mesh understanding, generation, and context-aware multi-turn editing inside a single Mixture-of-Transformers architecture.
READ POST →EVA01 — Unified Native 3D Understanding, Generation, and Multi-turn Editing
The technical report: how an understanding expert and a generation expert are coupled through shared global attention, plus the data pipeline and evaluation behind EVA01.
READ POST →PEGA: physics embedded generative architecture for real-time worlds
PEGA trains graphics engines and video models as one end-to-end system through a shared physics-embedded representation for persistent real-time worlds.
READ POST →PEGA-4world: explicit 3D memory + 2D video generation
An initial PEGA implementation using 3DGS as the intermediate representation — pairing explicit 3D memory with video generation to preserve geometry and sustain long-term scene coherence.
READ POST →
More from research
Papers
Peer-facing papers and preprints from Seele AI research.
Browse papers →Technical reports
In-depth reports on the architecture, methodology, and results behind our models.
Browse reports →Open source
Open models, datasets, and resources released around the EVA and PEGA lines.
Browse projects →The SeeleAgent system
How Seele02, EVA01, and the PEGA world model are orchestrated into one playable game.
Read more →