MoleculeDesk← Back to site

How to run Boltz-2 locally

MoleculeDesk can install and run Boltz-2 on an Apple Silicon Mac using its MPS runtime. This path has been verified with a real prediction. The Linux x64/NVIDIA CUDA path has also been verified with real installs and predictions on an NVIDIA RTX 3090 (see the benchmarks). Boltz-2 support is beta.

What you need

  • An Apple Silicon Mac (MPS), or a Linux x64 machine with an NVIDIA GPU (CUDA).
  • A Mac or Linux computer. Nothing else needs to be installed first: the MoleculeDesk installer sets up what it needs.
  • Disk space and a network connection for the model environment and checkpoints during installation.

Run moldesk doctor before installing. It reports your detected hardware and software compatibility. See Boltz-2 model requirements for the current Python, platform, accelerator, and file-format details.

Install and run a prediction

curl -fsSL https://moleculedesk.com/install.sh | sh
moldesk doctor
moldesk install boltz
git clone https://github.com/pooriyapfn/moldesk.git
moldesk run boltz moldesk/examples/boltz/protein.yaml

The example YAML input is a single-protein job sourced from Boltz's upstream examples. It uses msa: empty, so the prediction does not need an MSA search service. You can also provide your own Boltz .yaml, .yml, or .fasta file.

Find the outputs

The CLI reports the run status, run directory, and declared output file paths. Boltz-2 produces a predicted structure (.cif) and confidence data (.json); affinity output (.json) applies when the input requests it. See CLI usage for --output <dir> and machine-readable --json results.

Is this the right model?

Boltz-2 predicts biomolecular structures and can estimate binding affinity for suitable inputs. For structure-conditioned sequence design, see ProteinMPNN or LigandMPNN. For blind small-molecule docking, DiffDock-L is planned in MoleculeDesk and is not yet live-verified.

MoleculeDesk manages installation and execution; scientific interpretation and validation of predictions remain part of the research workflow. See the upstream Boltz project for model methods and input details.