Your phone fleet as a training cluster: Fractal slices models across Androids

What it does Fractal is a two-part ecosystem: FractalCore, a Python/Flask control plane that handles multi-tenant task dispatch, dataset segmentation, and FedAvg weight aggregation; and FractalAndroid, a native Kotlin app that does on-device TFLite training and memory-mapped inference. Phones poll the server for tasks, train locally on their own data bins, and upload only checkpoint deltas — raw data never leaves the device. The server credits nodes with tokenized rewards tracked in Firestore.

The interesting bit The second mode is the unusual one: for foundation models like Llama 3 8B, the server quantizes to INT4 (via torchao), traces the graph with torch.export, delegates to XNNPACK, and slices the result into ≤150MB layer_[N].pte partitions. Each Android node mmaps one slice and passes ≤4.1KB INT8 activation tensors to the next phone over a Wi-Fi socket — pipeline parallelism across a mesh of handsets, sized specifically to dodge Android’s Low Memory Killer.

Key highlights

• Federated learning with strict telemetry gates: training pauses unless battery ≥50%, charging, and ≤40°C
• Hard contracts everywhere — partition size, activation frame size, and single-use task_Id uploads are all enforced, with defined failure behavior
• Zero raw-data ingress: only weight checkpoints cross the wire
• Multi-tenant sandboxing at the directory level (data/tenants/{username}/)
• Stateless X-Auth-Token auth, deliberately isolated from browser cookies

Caveats

• Proprietary, source-available, non-commercial license — read it before building anything on top
• Requires physical Android devices (API 24+, ARM64 recommended), Firebase/Firestore, and a GCP project; not a laptop-friendly demo
• The README describes architecture in detail but offers no benchmarks on training throughput or inference latency

Verdict Worth a look if you’re into federated learning or genuinely novel edge-inference tricks — the phone-mesh pipeline parallelism is a rare design. Skip it if you need a permissively licensed, production-ready FL stack.