Odyssey

Odyssey is a family of multimodal protein language models developed by Anthrogen for protein sequence and structure generation, protein editing, and conditional design. Introduced in an October 2025 preprint, the family scales to over 102 billion parameters, with production models spanning roughly 1.2B to 102B, and was trained across an estimated 1.1 x 10^23 FLOPs — placing it among the largest biological models reported to date.

What distinguishes Odyssey is a combination of architectural choices aimed at scaling protein modeling efficiently. It represents proteins across multiple modalities — amino-acid sequence, 3D structure, semantic function descriptions, and orthologous-group metadata — fusing them in a single model. Atomic coordinates are tokenized with a finite scalar quantizer (FSQ), letting structure be modeled in the same discrete framework as sequence, and generation is cast as a discrete diffusion process in which sampling proceeds by iteratively unmasking tokens.

Odyssey is a proprietary system. Anthrogen has released a preprint describing the model but has not published training code or model weights, and the preprint is distributed under a no-derivatives, non-commercial license. The model is available only through Anthrogen rather than as an open checkpoint, which sets it apart from openly released protein models such as ESM or open backbone generators.

#Key Features

• Multimodal representation: Jointly models amino-acid sequence, 3D structure, functional/semantic descriptions, and orthologous-group metadata, enabling generation and editing conditioned on any of these signals.
• Consensus in place of attention: Replaces standard self-attention with a Consensus mechanism that propagates information through iterative local residue-level agreement and scales linearly, O(L), with sequence length, reported to improve training stability at scale.
• Discrete diffusion generation: Frames sequence and structure generation as a time-dependent unmasking procedure, allowing joint denoising of sequence and structure tokens.
• FSQ structure tokenization: Uses a finite scalar quantizer to discretize atomic coordinates, unifying structure and sequence within one token space and supporting strong structure-discretization performance.

#Technical Details

Odyssey couples an FSQ tokenizer for atomic coordinates with a transformer-style stack for multimodal representation learning, and trains via discrete diffusion. Its Consensus operator replaces quadratic self-attention with an iterative propagation scheme informed by local residue agreements, giving O(L) scaling in sequence length. Production models range from about 1.2 billion to over 102 billion parameters, trained across roughly 1.1 x 10^23 FLOPs, and Anthrogen reports the family was built by a small team with substantially lower cost and greater data efficiency than comparably sized efforts. The preprint reports strong results on protein generation and structure-discretization benchmarks, supported by theoretical analysis of the consensus mechanism.

#Applications

Odyssey targets protein engineers and drug-discovery teams who need to generate, edit, or conditionally design proteins guided by sequence, structure, or functional descriptions. Its editing and conditional-design capabilities suit workflows where a desired property or partial specification is used to steer generation toward candidate proteins. Because the model is proprietary and not distributed as an open checkpoint, access is mediated through Anthrogen rather than local deployment.

#Impact

Odyssey is notable for pushing protein language models toward frontier scale while introducing a linear-complexity consensus operator and a unified sequence-structure diffusion framework, and for reporting these results from a small team at comparatively low cost. Its practical influence on the field is constrained by its closed status: with no released code or weights and a non-commercial, no-derivatives preprint license, external groups cannot yet reproduce or directly build on the model, and its generative claims await independent and experimental validation.