Skip to main content

FAQ

General​

What is Octomil? Octomil is a platform for running ML models on edge devices -- local inference, on-device training, and federated learning across iOS, Android, and desktop.

How is Octomil different from Flower? Flower is a research framework for federated learning. Octomil is a production platform with a dashboard, mobile-first SDKs, model rollouts, A/B testing, and compliance presets. Octomil's API is designed to require 5 lines of code, not 500.

What model formats are supported? PyTorch, ONNX, CoreML, and TFLite. Octomil handles conversion: PyTorch -> ONNX -> CoreML (iOS) / TFLite (Android). See Supported Models.

Setup & Configuration​

How do I get started? Follow the Quickstart. You will have a model running locally or on a device in under 10 minutes.

Do I need a cloud account to use Octomil? No. octomil serve <model> runs inference locally with no account needed. Cloud features (device fleet management, federated training, rollouts) require an Octomil account.

How do I configure compliance presets? Run octomil init "Your Org" --compliance hipaa --region us during setup, or apply a preset later via the dashboard. See Compliance.

Training & Federated Learning​

Does my data ever leave the device? No. Training data stays on-device. Only model weight deltas (numerical arrays representing how the model changed) are transmitted to the server.

What aggregation strategies are available? FedAvg, FedProx, FedOpt, FedAdam, Krum, MultiKrum, FedMedian, FedTrimmedAvg, and SCAFFOLD. See Advanced FL Configuration and Advanced FL Strategies.

How many devices do I need for federated training? A minimum of 3 devices per round is recommended. For Byzantine-robust strategies, ensure at least 3x your estimated number of unreliable clients.

What happens if a device drops out mid-round? The server waits for the configured timeout, then aggregates whatever updates have been submitted. Enable async mode for fleets with high dropout rates.

Privacy & Security​

Is Octomil HIPAA compliant? Octomil is architecturally incapable of violating HIPAA's core data protection requirements because it never handles PHI. BAA execution is available on the Enterprise tier. See Compliance.

What is differential privacy and do I need it? Differential privacy adds calibrated noise to model updates so individual training examples cannot be reverse-engineered. Enable it for sensitive workloads. See Privacy Guide.

How does secure aggregation work? With secure aggregation enabled, the server can only see the sum of client updates, not individual updates. This prevents the server from inferring information about any single device's data.

SDKs​

Which platforms are supported? Python, iOS (Swift + CoreML), Android (Kotlin + TFLite), and Browser (WebGPU + WASM). See the SDK docs.

Can I use the OpenAI client library with Octomil? Yes. Octomil's local inference server (octomil serve) exposes an OpenAI-compatible API at http://localhost:8080/v1. Use any OpenAI client library with base_url pointed to your local server.

Billing & Plans​

Is there a free tier? Yes. The free tier includes local inference with no limits. Cloud features (device management, federated training) include a generous free allocation.

How do I upgrade to Enterprise? Contact team@octomil.com for Enterprise pricing, BAA execution, dedicated support, and custom SLAs.