NimbusNimbus
Back to Blog

Nimbus Studio: February 2026 Product Updates

Published March 2026

February was a major release month for Nimbus Studio. We shipped the Nimbus Studio desktop app, cloud GPU offloading, and the ZUNA foundation model for EEG preprocessing. Here’s what’s new.

Desktop App: Run Nimbus Studio on Your Machine

Screenshot_2026-03-07_at_22.10.20.png

The Nimbus Studio desktop app is now available as a standalone Electron application. Download and run the full pipeline builder with an embedded backend—no separate setup required.

Highlights:

  • Download and run: Install the app and launch with a single click
  • OAuth via system browser: Sign in with Clerk (GitHub, Google, email) in your browser; after OAuth, the app reopens automatically
  • Public dataset downloads: MOABB and other public datasets download on demand when needed

The desktop app uses the same pipeline builder as the web app, so your workflows stay consistent whether you use the browser or the app.

GPU Offload: Cloud GPUs for Heavy Compute

Heavy compute nodes can now be offloaded to cloud GPUs via Modal. This lets you use GPU acceleration when running on CPU-only hosts or when on-device GPUs aren’t available.

Supported nodes:

  • ZUNA (diffusion model): A100-40GB
  • REVE (transformer): L40S
  • EEGNet (CNN training): T4

Features:

  • Per-node toggle: Enable “Cloud GPU” in the pipeline editor for each supported node
  • Baked model weights: ZUNA (~1.5GB) and REVE (~400MB) weights are baked into the cloud image at deploy time, so there’s no download delay at runtime
  • Retry logic: Transient failures (network, timeouts) are retried with exponential backoff
  • Graceful fallback: If Modal is unavailable or fails, execution falls back to on-host CPU/GPU

GPU offload is available when Nimbus Studio is deployed with the Modal integration enabled.

ZUNA Preprocessing Node: Foundation Model for EEG

zuna.jpeg

ZUNA is a new optional preprocessing node for EEG denoising, channel reconstruction, and upsampling, powered by the ZUNA foundation model.

Capabilities:

  • EEG denoising via diffusion
  • Channel reconstruction and upsampling
  • Batch-only (not available in streaming)
  • Cloud GPU recommended (5–10 min runtime)

Configuration options: Cloud GPU toggle, diffusion steps, notch/highpass filters, average reference, montage, upsample channels, and bad channel handling.

ZUNA is available as a preprocessing node in Nimbus Studio. The first run downloads model weights (~1.5 GB) from HuggingFace. Outputs are imputed—use with care in clinical settings.

Getting Started