EVA01: Unified Native 3D Understanding and Generation via Mixture-of-Transformers

This paper addresses the challenge of incorporating 3D mesh directly into the sequence of Multimodal Large Language Models (MLLMs), rather than treating it as an external output. Diffusion-based large reconstruction models decouple semantic understanding from geometric reasoning, operating as stateless reconstructors conditioned on dense 2D pixel priors. Recent MLLM-based methods treat the 3D modality as an external output rather than an integral component of the multimodal sequence, making incremental adaptations without systematic analysis of how geometric manifolds align with MLLM feature spaces. We introduce EVA01, a framework that extends the modality boundary of MLLMs to incorporate 3D mesh understanding, generation, and context-aware editing, processing 3D tokens directly within the shared sequence rather than through rendered 2D intermediaries. Built upon a Mixture-of-Transformers (MoT) architecture, EVA01 decouples the model into a pre-trained Understanding Expert ( E und ) and a structurally mirrored Generation Expert ( E gen ), coupled through shared global self-attention with hard modality routing. This design aligns the semantic latent space of the MLLM backbone with the geometric manifold, enabling direct transfer of multimodal priors without intermediate 2D representations. Results show that EVA01 achieves state-of-the-art text-to-3D generation fidelity and unlocks robust long-context multi-turn geometric editing with identity preservation—a capability fundamentally inaccessible to stateless reconstruction pipelines. Our findings further offer architectural insights for integrating 2D foundation models with 3D tasks, informing the design of multimodal systems that operate directly on 3D. Project page: https://github.com/SeeleAI/OpenEVA.