ByteDance releases Protenix-v1, an open-source model for predicting biomolecular structures

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

What Protenix-v1 is

MarkTechPost describes Protenix-v1 as a "comprehensive reproduction of AlphaFold3": not a stripped-down analogue, 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). That means the system can be run and fine-tuned independently, without going through a closed API or a paid service.

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

The Apache 2.0 license, under which the model was released, is one of the most permissive open licenses. It allows commercial use, modification and distribution of derivative versions without any obligation to disclose their source code. For labs and startups, this means Protenix-v1 can be built into their own products and pipelines without the legal restrictions typical of stricter copyleft licenses.

How the model is compared with AlphaFold3

The key question MarkTechPost poses is how close an open model can get to AlphaFold3's accuracy if three parameters are equalized: training data, model scale and inference budget. The framing matters. It treats Protenix-v1 not as a standalone product with abstract claims about quality, but as a model that can be evaluated under clear, comparable conditions against a specific benchmark target.

"How close can an open model get to AlphaFold3 accuracy when training data, model scale and inference budget are matched?"

This is how MarkTechPost phrases the central question of its piece on Protenix-v1.

The wording itself points to the methodology used to describe the model: the comparison relies not on absolute benchmark figures but on equal starting conditions, meaning the same amount of training data, a similar parameter scale and the same inference budget. This makes ByteDance's claim easier to verify than marketing comparisons that give no conditions.

Which structures the model predicts

Protenix-v1 targets several classes of biomolecules at once: proteins, DNA, RNA and ligands. That is the same multitask coverage AlphaFold3 is known for. It sets the model apart from narrowly specialized systems that handle only one type of structure: the developers state explicitly that the goal is to cover all the main classes AF3 works with within a single architecture.

Publishing the code and weights openly widens the circle of people who can work with a model of this caliber. Labs and research groups without access to closed commercial systems can reproduce experiments and build Protenix-v1 directly into their own pipelines, rather than going through a restricted external API.

What this means

An open model that aims for AlphaFold3-level accuracy across several classes of biomolecules, and ships with full code, weights and the permissive Apache 2.0 license, lowers the barrier to entry into structural biology for teams without access to closed systems. It also lets others check the claimed accuracy independently, rather than taking the developer's word for it.