ByteDance Released Protenix-v1: Open-Source Model for Biomolecular Structure Prediction

ByteDance on February 8, 2026 presented Protenix-v1 — an open model for predicting biomolecular structure, which reproduces the architecture and capabilities of AlphaFold3 (AF3) and is distributed with source code and model weights under the Apache 2.0 license.

What is Protenix-v1

Protenix-v1 is, according to MarkTechPost, a "comprehensive reproduction of AlphaFold3": not a stripped-down analog, but a model aimed at the same level of accuracy under comparable conditions. ByteDance published not only the code, but also the model weights (model parameters) themselves — which means the system can be run and fine-tuned independently, without recourse to a closed API or paid service.

• Model developer — ByteDance
• Release name — Protenix-v1
• Stated goal — accuracy at the level of AlphaFold3 (AF3)
• Code and weights published under Apache 2.0 license
• Target structure types — proteins, DNA, RNA and ligands
• MarkTechPost published material about the release on February 8, 2026

The Apache 2.0 license, under which the model was released, is one of the most permissive among open licenses: it allows commercial use, modification and distribution of derivative versions without the obligation to disclose their source code. For laboratories and startups, this means that Protenix-v1 can be embedded in their own products and pipelines without the legal restrictions characteristic of stricter copyleft licenses.

How the model compares to AlphaFold3

The key question posed by MarkTechPost is: how close can an open model get to AlphaFold3 accuracy if three parameters are equalized: training data, model scale and inference budget. This framing is important: it turns Protenix-v1 not into a standalone product with abstract quality claims, but into a model that can be evaluated against clear, comparable criteria against a specific benchmark.

"How close can an open model get to

AlphaFold3 accuracy if training data, model scale, and inference budget are equalized?" — this is how MarkTechPost frames the central question of the Protenix-v1 material.

The formulation of the question itself indicates the methodology used in describing the model: the comparison is not based on absolute benchmark numbers, but on equal inputs — that is, with the same amount of training data, similar model scale and equal inference budget. This makes ByteDance's claim more verifiable than marketing comparisons without specified conditions.

What structures the model predicts

Protenix-v1 targets several classes of biomolecules at once — proteins, DNA, RNA and ligands, which is the same versatility for which AlphaFold3 is known. This distinguishes the model from narrowly specialized systems that work with only one type of structure: the developers explicitly state that the goal is to cover all major classes that AF3 works with, within a single architecture.

The open publication of code and weights expands the circle of those who can work with a model of this level. Laboratories and research groups without access to closed commercial systems gain the ability to reproduce experiments and embed Protenix-v1 in their own pipelines directly, rather than through a limited external API.

What this means

An open model that aims for AlphaFold3 accuracy level across multiple biomolecule classes and is distributed with complete code, weights and a permissive Apache 2.0 license lowers the barrier to entry into structural biology for teams without access to closed systems and allows the claimed accuracy to be verified independently, rather than taken at the developer's word.