Skip to content
FyreFlySystems
Abstract 3D render of a head in profile formed from tangled orange, violet and pink tubes that unravel into separate strands
Go to Phonos ↗

Phonos

Structure prediction, docking, protein design and molecule generation for research teams — in the browser, without your own GPUs.

To Predict

Google DeepMind / Pexels ↗

Run the models, not the infrastructure

Predict structures and complexes, dock ligands, design proteins and generate molecules and DNA — then inspect the results in 3D and pass them on to the next model.

Abstract 3D render of yellow, orange and coral-pink organic forms linked by thin annotation lines Google DeepMind / Pexels ↗

For computational biologists and the lab teams they work with: published models behind one interface, with your data, results and colleagues in the same project.

  • Bring your data. Upload sequences, structures and ligands, or import them from the Protein Data Bank, the AlphaFold Database, UniProt and PubChem.
  • Run and chain models. Start a single job, or a pipeline such as fold, then dock — Phonos checks that every stage gets the input it needs.
  • Inspect and share. Open results in the 3D viewer, analyse them in a notebook, and work on them together with your project members.
  • Script it. The same models, jobs, projects and files over a REST API with API keys.

Phonos brings the models, the files and the people of a research project into one place. You work in projects: upload your data or import it from a public database, run a model, open the result in the 3D viewer and pass it on to the next model — without a terminal, a Docker stack or DevOps.

Phonos is offered as a Research Preview: a testing phase in which features, compute and storage may change. It is a research tool, not a medical device.

Runs in
Your browser
Languages
English, German
Status
Research Preview
Access
Trusted Tester Program
Address
fyreflysystems.de/phonos ↗
See the models

Eight models, one interface

Published models from research groups and NVIDIA. In the Research Preview they are served through NVIDIA BioNeMo NIM, for testing; paid plans run them on machines operated by FyreFly Systems. Each model page lists what it takes, what it returns, the limits per job and how to read its confidence.

Boltz-2

Structures of complexes — proteins, nucleic acids and small molecules — plus the binding affinity of a ligand.

Structure · MIT Jameel Clinic Inputs and limits

OpenFold3

Structures of protein, DNA or RNA chains, optionally with a ligand — with or without an alignment.

Structure · OpenFold Consortium Inputs and limits

DiffDock

Docks a small molecule into a protein and ranks the poses by confidence.

Docking · MIT CSAIL Inputs and limits

RFdiffusion

Generates new protein backbones: binders against a target, scaffolds around a motif or proteins from scratch.

Protein design · IPD, University of Washington Inputs and limits

ProteinMPNN

Designs amino-acid sequences predicted to fold into a given backbone.

Protein design · IPD, University of Washington Inputs and limits

MolMIM

Generates small molecules from a seed and optimises them for QED or logP.

Small molecules · NVIDIA Inputs and limits

GenMol

Completes a SMILES seed or a masked fragment into new, valid drug-like molecules.

Small molecules · NVIDIA Inputs and limits

Evo 2

A genomic language model with 40 billion parameters; on Phonos it generates DNA that continues your sequence.

DNA · Arc Institute Inputs and limits

Retired: ESMFold, AlphaFold2 and AlphaFold2-Multimer. NVIDIA closed their hosted endpoints in August and September 2026, so they no longer run on Phonos; for structures and complexes use Boltz-2 or OpenFold3. Details →

All models, compared in one table

Four pillars

Phonos was built to cut through the complexity of biotech computation. Everything we do rests on these four.

Accessibility
Every model on Phonos is reachable in a few clicks — no terminal, no Docker stack, no DevOps. The interface is in English and German.
Cost-effectiveness
Jobs count per run, not per GPU hour, and a pipeline counts once for all of its stages.
Where your data goes
Your account, projects, files and results are stored on a server FyreFly runs in Germany; requests reach it through Cloudflare's network. Models served through NVIDIA NIM run on NVIDIA DGX Cloud: the input you submit to them — sequences, structures, ligands, parameters — is transmitted to NVIDIA and may be processed in the United States, under EU Standard Contractual Clauses and the EU–US Data Privacy Framework. Phonos marks these models in the interface. Privacy policy →
Collaboration
Project-oriented workflows keep teams organised. Shared workspaces, role-based access and real-time visibility turn isolated research into coordinated progress.

How it works

A signed job queue with live results, pipelines that chain models, projects with roles and a notebook workbench.

From queue to browser

Every job takes the same path, and every result is checked before it reaches you.

How a job runs

Pipelines

Chain models from different stages into one run.

Chain models

Projects, roles and live collaboration

Research on Phonos happens in projects. Each member has one of three roles.

Projects and roles

More in the workbench

Notebook IDE, public databases, REST API, model chaining, security and two languages.

See the workbench

Getting access

Phonos is a Research Preview. Early access runs through our Trusted Tester Program.

Abstract collage of pink and lime-green coral-like forms over an aerial view of braided river channels, marked with thin square frames Google DeepMind / Pexels ↗

  1. Create an account

    Sign up at fyreflysystems.de/phonos.

  2. Apply as a Trusted Tester

    In Phonos, tell us in a few lines about your research, your institution and what you want to use Phonos for.

  3. We review it

    We read every application and let you know in Phonos and by email. Trusted Testers move on to regular access later.

Open Phonos ↗ Questions first? Write to us