Odyssey: Anthrogen's 102B Protein Model Replaces Attention with Consensus and Uses Discrete Diffusion

A multimodal model for joint sequence and structure

Anthrogen revealed Odyssey, a family of multimodal protein language models designed for sequence and structure generation, protein editing and conditional design. Production-scale variants span from 1.2B to 102B parameters. The team positions Odyssey as a practical system for real protein design workloads and reports early API access for users.

Input representation and FSQ structure tokens

Odyssey fuses sequence tokens, discrete structure tokens and lightweight functional cues into a shared latent representation. To convert 3D geometry into compact tokens the model uses a finite scalar quantizer (FSQ) — an ‘‘alphabet of shapes’’ that lets the network treat structure alongside amino acid sequence. Functional cues can include domain tags, secondary structure hints, ortholog group labels or short text descriptors, giving Odyssey both local sequence context and long-range geometric relations.

Consensus propagation replaces global self-attention

Rather than global self-attention, Odyssey uses a new propagation rule called Consensus. Consensus performs iterative, locality-aware updates on a sparse contact or sequence graph: neighborhoods first agree locally, and that agreement then spreads outward across the chain and contacts. Anthrogen reports that Consensus scales as O(L) with sequence length L, compared with the O(L²) cost of self-attention, reducing compute for long proteins and multi-domain constructs. They also report improved robustness to learning rate choices at larger scales, lowering brittle runs and restarts.

Training with discrete diffusion

Odyssey trains on sequence and structure tokens using a discrete diffusion objective. The forward process applies masking noise that mimics mutation, while the reverse-time denoiser learns to reconstruct sequence and coordinates jointly. At inference the reverse denoising process supports conditional generation and editing: users can fix a scaffold, hold a motif, mask a loop, add a functional tag and let the model complete the rest while keeping sequence and structure synchronized.

Performance, data efficiency and practical implications

In matched comparisons Anthrogen reports that diffusion training outperforms masked language modeling during evaluation, with lower training perplexities versus complex masking and comparable or better perplexities versus simple masking. They also note that Odyssey achieves strong performance with roughly 10x less training data than competing models, addressing data scarcity in protein modeling. The system targets multi-objective design including potency, specificity, stability and manufacturability, and its O(L) scaling for Consensus makes long proteins and multi-domain designs more tractable at large parameter counts.

What this means for protein design workflows

Odyssey operationalizes joint sequence and structure modeling via FSQ, Consensus propagation and discrete diffusion. The combination enables conditional design and targeted editing under realistic constraints and aims to reduce compute and data requirements for large-scale protein engineering tasks. Interested readers can consult the Anthrogen paper for technical details and check the project’s repositories and resources for code, tutorials and community channels.