Run Boltz-2 for protein structure and binding-affinity prediction
Boltz-2 is an open-source biomolecular model that jointly predicts 3D structure and binding affinity, released under the MIT license by the Boltz team. MoleculeDesk manages Boltz-2's Python environment, checkpoints, and commands so you can use it without configuring CUDA, PyTorch, or dependency versions by hand.
Status: beta. moldesk install boltz and moldesk run boltz work end-to-end on Apple Silicon (Darwin arm64), verified with a real install and prediction using PyTorch MPS. The Linux x64/NVIDIA CUDA path is now verified too: a real install and prediction ran on an NVIDIA RTX 3090 (see the benchmarks below).
What Boltz-2 is used for
Boltz-2 is designed for:
- Protein structure prediction — predicting 3D structure from sequence.
- Binding-affinity prediction — estimating how strongly a ligand or biomolecule binds a target, jointly with structure.
- Biomolecular complex modeling — modeling multi-chain complexes, not just single proteins.
These are Boltz-2's upstream design goals; MoleculeDesk's job is to make installing and running it painless, not to change what it can predict.
Requirements
| Model version | 2.2.1 |
| License | MIT |
| Python | 3.11, via uv |
| Platforms | darwin-arm64 (verified), linux-x64 (verified on RTX 3090) |
| Accelerator | Apple Silicon MPS (verified) or NVIDIA GPU/CUDA (verified); CPU also accepted |
| Input formats | .yaml, .yml, .fasta |
| Output | structure (.cif), confidence (.json), optional affinity (.json) |
Boltz-2 is a genuine dual-runtime integration: Linux x64 installs the
official jwohlwend/boltz with its
CUDA extra, while Darwin arm64 installs the
boltz-community
fork, which adds PyTorch MPS support. MoleculeDesk selects the right one
automatically based on your platform.
Boltz-2 benchmarks: speed, VRAM and cost per prediction
On an NVIDIA GeForce RTX 3090, Boltz-2 takes about 85 s per prediction once warm (2.3 min on the first run), roughly 42 predictions per GPU-hour, or about $0.012 per prediction.
| Metric | NVIDIA GeForce RTX 3090 |
|---|---|
| Time per run, first (cold) | 2.3 min |
| Time per run, warm | 85 s |
| Predictions per GPU-hour | 42 |
| Cost per run, first (cold) | $0.019 |
| Cost per run, warm | $0.012 |
| Peak GPU memory | 2.6 GiB |
| Peak GPU utilization | 60 % |
| Peak GPU power | 163 W |
| Install time | — |
| Disk per install | 17 GiB |
Workload: Single protein chain, 115 residues, no MSA (examples/boltz/protein.yaml), 1 sample. Median of 3 runs.
Machine: RunPod GPU pod, Linux x64, 125 GiB RAM; AMD EPYC 7H12 (32 vCPU allocated); driver 580.126.20.
Cost: at $0.50/GPU-hour (RunPod Secure Cloud, EU-CZ-1, compute only; storage adds about $0.02/hr); excludes storage and idle time.
Warm time is the mean of runs 2 and 3 (86 s and 84 s). Very first run ever also downloads the CCD data (~198 s total).
Source
MoleculeDesk pins each platform's Boltz-2 distribution to a specific
upstream commit rather than tracking a moving branch, so installs are
reproducible. See the upstream Boltz repository
(Linux/CUDA) and boltz-community
(Darwin/MPS) for full details on the model itself.
FAQ
Does moldesk install boltz work today?
Yes. Verified with a real install and prediction on Apple Silicon (MPS) and on Linux with an NVIDIA RTX 3090 (CUDA).
What license is Boltz-2 under? MIT. MoleculeDesk doesn't redistribute Boltz-2's weights; installing it fetches them under Boltz's own terms.
What models are available today? Boltz-2 (beta, Apple Silicon verified), ProteinMPNN, and LigandMPNN all have working (beta) installation today.