Questions
Access, models, where your data is stored, what reaches NVIDIA, scripting, clinical use and retired models.
Google DeepMind / Pexels ↗
How do I get access?
Phonos is a Research Preview, and early access runs through the Trusted Tester Program. Create an account at fyreflysystems.de/phonos ↗, then apply in Phonos with a few lines about your research, your institution and your use case. We review every application and tell you the result in Phonos and by email.
Which models can I run?
Eight: Boltz-2 and OpenFold3 for structures and complexes, DiffDock for docking, RFdiffusion and ProteinMPNN for protein design, MolMIM and GenMol for small molecules, and Evo 2 for DNA. The models page lists inputs, outputs and limits for each.
Where is my data stored?
Accounts, projects, files and results are stored on a server FyreFly runs in Germany; requests reach it through Cloudflare's network. Inputs to models served through NVIDIA go to NVIDIA (see the next question). The privacy policy lists every processor and where it is located.
Which models send my input to NVIDIA?
The models served through NVIDIA BioNeMo NIM — the models page names the server for each one, and Phonos marks them in the interface. The input you submit to them is transmitted to NVIDIA and may be processed in the United States. Do not submit confidential, proprietary or personal data to these models that you are not comfortable sharing with NVIDIA under its terms.
Can I use Phonos from a script?
Yes. Create an API key in Phonos and send it as a Bearer token. The REST API covers models, jobs, projects and files, accepts jobs in batches, and returns the same compatibility table the browser uses to chain models.
Can I use results for clinical decisions?
No. Phonos is a research and educational tool, not a medical device. Its outputs must not be used for clinical, diagnostic or therapeutic decisions, and they should be verified independently before you rely on them.
What happens when a provider retires a model?
The model is marked as retired and no longer runs; jobs still waiting for it are stopped with a message. NVIDIA retired the hosted ESMFold, AlphaFold2 and AlphaFold2-Multimer endpoints in 2026 — use Boltz-2 or OpenFold3 instead. The models page lists them.