Boltz

Learning Objectives

• Understand what biomolecular foundation models like Boltz-2 and BoltzGen predict and design
• Tell apart the open-source Boltz models from the paid, hosted Boltz pipelines
• Read Boltz's published results for what they are: early experimental evidence, mostly from the company itself

What Is Boltz?

Boltz is a family of AI models for drug discovery, and the company that builds them. Boltz is a public benefit corporation in Cambridge, Massachusetts, founded by the MIT researchers who released the open-source Boltz models. The models do three related jobs. They predict the three-dimensional shape of proteins and the molecules bound to them, they estimate how tightly a small molecule binds its target, and they design new proteins that stick to a chosen target.

The research models are open. Boltz-1 arrived in November 2024, Boltz-2 in June 2025, developed with Recursion and able to predict binding strength with accuracy the team says approaches physics-based simulation at more than a thousand times the speed, and BoltzGen in October 2025, which designs new protein binders. All three are released under the MIT license, weights included, so anyone can download and run them, including for commercial work.

The company's paid products sit on top. In June 2026 Boltz launched BoltzMol-1, a pipeline for finding small-molecule starting points, and BoltzProt-1, a pipeline for designing nanobodies and other protein binders. These are proprietary, available only through the hosted Boltz Lab platform and the Boltz API. Boltz says its models have been downloaded more than five million times and are used by more than a hundred thousand scientists.

Pricing

The open models cost nothing to download; you pay only for your own computing. The hosted platform has a free plan with a monthly allowance, usage-based pricing beyond it, and a custom enterprise plan.

Boltz says customers own what they create and that it does not train its models on customer materials without permission.

Core Features

Structure Prediction

Boltz models predict how a protein folds and how it sits against other molecules, including drug candidates, DNA and RNA, in the lineage of AlphaFold.

Binding Affinity

Boltz-2 adds an estimate of how strongly a small molecule binds its target, the number medicinal chemists care about when ranking candidates.

Binder Design

BoltzGen generates new proteins and peptides designed to bind a chosen target. The paid BoltzProt-1 pipeline extends this to nanobodies, and Boltz reports it nearly tripled BoltzGen's hit rate in its own tests.

Small-Molecule Hit Discovery

BoltzMol-1 ranks existing purchasable compounds or generates new ones and proposes a short list for lab testing. Boltz reported confirmed hits on six of ten difficult targets while testing between 28 and 51 compounds per target.

Agents and Integrations

Boltz Lab wraps the models in agents for small-molecule and protein design. Since June 2026 the Boltz API has been available inside Anthropic's Claude products, so an AI agent can call the models directly as part of a longer workflow.

Custom Models for Pharma

Enterprise customers can have models fine-tuned on their own data. Pfizer signed the first such collaboration in January 2026, followed by Takeda, GSK and the flavor and fragrance company dsm-firmenich.

Strengths

• Genuinely open research models — MIT license on code and weights, with no revenue or territory limits
• Widely used — a large academic and industry user base, according to the company
• Low entry cost — a free hosted allowance, and predictions priced in cents
• Covers both small molecules and proteins — structure, affinity and design in one family
• Big-pharma collaborations — Pfizer, Takeda and GSK have signed on to deploy the models

Limitations and Considerations

• The newest models are closed — BoltzMol-1 and BoltzProt-1 are hosted only, unlike the open research models
• Hosted terms carry restrictions — internal business use only, no users in US-embargoed countries, and no training competing models on the outputs
• Most evidence is self-reported — the BoltzMol-1 results are a company technical report and preprint, not yet peer reviewed
• Hits are starting points — the Werner helicase compounds Boltz reported in September 2026 were active in the micromolar range, far from a drug
• Self-hosting needs expertise — running the open models well takes GPUs and computational chemistry skill
• Not a clinical tool — no regulatory standing, and designs must be synthesized and tested in a lab

Best Use Cases

How It Compares

Boltz sits between two approaches in this category. Chai Discovery licenses its newest antibody-design models to drug companies rather than releasing them, and Isomorphic Labs keeps its drug design engine private. Latent Labs, Profluent and Cradle focus on protein design. Boltz is open at the research layer and paid at the product layer.

Key Takeaways

• Boltz builds biomolecular AI models for structure prediction, binding affinity and protein and small-molecule design
• Boltz-1, Boltz-2 and BoltzGen are MIT-licensed, weights included, with no revenue, territory or training limits
• The newer BoltzMol-1 and BoltzProt-1 pipelines are proprietary and reached through Boltz Lab or the API, with 200 free predictions a month and usage pricing beyond
• Pfizer, Takeda, GSK and dsm-firmenich have signed collaborations, with no financial terms disclosed
• Its published results are promising early evidence, largely from the company itself, and every design still needs testing in a lab