What's New in Nimbus: July 2026 SDK Update

July was a big month for the Nimbus SDK.

We shipped v0.5.0, published a preprint explaining our approach to BCI personalization, and started the road to 0.6 with a focus on easier adaptation, clearer confidence signals, and more reliable performance across users and sessions. If you’re newer to the SDK’s positioning as a decision layer, start with Why Nimbus SDK? Beyond scikit-learn and pyRiemann for BCI.

This month’s theme is simple: make BCI personalization easier to integrate, easier to measure, and easier to trust.

Nimbus SDK v0.5.0 is available on PyPI now.

1. The preprint: one personalization layer for many EEG models

We posted a preprint that formalizes the main idea behind July’s SDK work:

Universal BCI Personalization: One API for Frozen EEG Trunks and Foundation Models Sergey Musienko — arXiv:2607.22397 Submitted 24 Jul 2026 https://arxiv.org/abs/2607.22397

The problem is that EEG models are multiplying. Teams now have classical deep-learning models, lightweight production models, and emerging foundation models. But if every model needs its own personalization system, product integration becomes slow and hard to maintain.

The SDK’s answer is a shared personalization layer. A BCI team can connect an EEG model once, then use the same confidence, calibration, and adaptation workflow around it.

That matters because it separates model choice from product behavior. You can test different models while keeping the same user experience: confidence-aware predictions, smarter calibration, and better decisions about when the system should act or abstain.

The paper also supports a practical product point: in many cases, lightweight calibration is enough. You do not always need a full model retrain to improve user-specific performance.

2. Shipped in v0.5.0: personalization becomes a first-class SDK workflow

v0.5.0 is the release where personalization becomes a first-class SDK feature.

At a high level, the SDK can now take an existing EEG model and turn its output into a more useful BCI decision signal: not just “which class won?”, but “how confident are we, and what should we do next?”

What this means in practice:

• Teams can reuse existing EEG models. You do not need to rebuild the full stack every time you test a new model family.
• The output is more useful for real products. Applications can use confidence and uncertainty, not just raw predictions.
• Calibration can be shorter and smarter. Instead of collecting a fixed amount of data every time, the system can stop when the signal is good enough.
• The SDK can guide the next step. When confidence is low, the system can help decide whether to keep running, collect more data, or adapt.
• Foundation-model workflows are becoming easier to test. July’s release moves the SDK closer to a common personalization layer for both classical EEG models and newer foundation-model encoders.

In other words: 0.5.0 moves Nimbus SDK from “a set of Bayesian classifiers for EEG” toward “a personalization layer for real BCI products.”

3. In development next: a clearer personalization workflow

After 0.5.0, most July engineering goes into the future release track.

The goal is to make personalization easier to understand and compare. Instead of treating adaptation as a black box, the SDK is moving toward a clearer workflow:

1. Read the EEG signal with a model.
2. Adjust for the current user or session when needed.
3. Make a confidence-aware decision.

That separation helps answer practical questions:

• Is the current model good enough for this user?
• Do we need more calibration data?
• Is a lightweight adjustment enough?
• Or do we need a stronger adaptation step?

For product teams, this is the difference between “the model is uncertain” and “we know what to try next.”

4. Adaptation becomes measurable

One of the most important 0.6 improvements is that adaptation cost is becoming visible.

Instead of only asking “did accuracy improve?”, teams can also ask:

• How long did adaptation take?
• Was it lightweight enough for the product flow?
• Is the improvement worth the extra calibration time?
• Could this run on the target device or session setup?

This is especially important for BCI products. Users do not want long calibration sessions, and hardware teams often need predictable performance on constrained devices.

The SDK is moving toward making those tradeoffs clear, so teams can choose the simplest adaptation that works.

5. Better performance across users and sessions

The SDK is also improving how it handles differences between users, sessions, devices, and recording conditions.

This matters because EEG changes. A model that works well in one session may see a slightly different signal distribution in another. The goal is to make those shifts easier to handle without forcing teams into a heavy retraining workflow.

The practical product point: not every personalization problem requires retraining a neural network. Sometimes the right move is a simpler alignment step that makes today’s signal look more like the signal the model expects.

6. Upgrade notes

If you are upgrading from an earlier SDK version, check these areas first:

The main user-facing note: if you used saved alignment profiles from an older version, re-fit them after upgrading.

7. Docs and examples

Two onboarding pieces landed alongside the SDK work:

• Full SDK documentation: docs.nimbusbci.com
• 5-minute quickstart — from install to a confidence-aware brain-state prediction.
• End-to-end real-EEG example — a full motor-imagery workflow with rejection and ITR.

Upgrade

For the shipped release:

pip install -U nimbus-bci==0.5.0

What’s next

The direction is clear: Nimbus SDK is becoming the personalization and uncertainty layer for EEG systems, not just a collection of classifiers. The goal is to make adaptation decisions explicit, measurable, and cheap enough to use in real BCI products.

Links

• Preprint: https://arxiv.org/abs/2607.22397
• Docs: docs.nimbusbci.com
• PyPI: nimbus-bci