AMix-2: Establishing Protein as a Native Modality in Large Language Models

Researchers introduce AMix-2, a protein-text foundation model that treats protein sequences as a native modality in large language models alongside natural language. The model uses a novel block-wise diffusion approach instead of traditional left-to-right generation, paired with a new ProteinArena benchmark for evaluating protein AI systems.

AMix-2 represents a meaningful advancement in multimodal AI by establishing proteins as a first-class citizen within foundation models rather than treating them as specialized downstream tasks. The research addresses a fundamental limitation in current AI systems: protein understanding and design typically require separate, task-specific models. By unifying these capabilities in a single foundation model with shared token space between language and protein sequences, the approach enables more efficient biological reasoning and conditional generation.

The block-wise diffusion language modeling backbone demonstrates sophisticated technical thinking about protein sequence generation. Rather than strictly sequential left-to-right generation common in autoregressive language models, this architecture combines causal block-level generation with bidirectional context and iterative refinement within blocks. This better reflects how proteins actually fold and function—not as purely linear sequences but as complex 3D structures where distant regions interact.

The introduction of ProteinArena as a comprehensive benchmark signals maturation in how the field evaluates protein AI systems. Time-aware and homology-aware evaluation protocols address realistic generalization challenges, moving beyond isolated benchmarks that may not capture true model capability. AMix-2's competitive performance against specialized protein models while outperforming frontier LLMs suggests the unified foundation model approach offers genuine advantages.

For the broader AI ecosystem, this demonstrates that foundation models can effectively incorporate domain-specific modalities when properly architected. The open release of both AMix-2 and ProteinArena will likely accelerate protein AI research. However, real-world impact depends on downstream adoption in drug discovery, synthetic biology, and protein engineering applications where empirical validation matters more than benchmark performance.

• →AMix-2 unifies protein understanding and design in a single foundation model using shared token space between language and protein sequences
• →Block-wise diffusion architecture outperforms autoregressive generation for protein sequences, indicating flexible generation order improves results
• →ProteinArena benchmark includes time-aware and homology-aware protocols for realistic evaluation across understanding and design tasks
• →Model demonstrates competitive performance to specialized protein tools while outperforming general-purpose frontier LLMs on protein tasks
• →Open release of model and benchmark aims to accelerate research in protein foundation models as native multimodal AI capability