From a Pi on your desk to your first thousand devices.

Prototype → first deployment

Start on a Pi 5 with Hailo-10H AI HAT+ 2 ($130) or Jetson Orin Nano Super ($249). Furcate SDK ships a working example for Qwen3-VL or Pixtral inference, OpenVLA for cobot tasks, LiteRT for vision, TFLM for sensor anomaly. Iterate on the device; deploy to a fleet when it's ready.

Bring your own model

ONNX, GGUF, GGML, SavedModel, TorchScript — Furcate's quantization pipeline takes a foundation-model checkpoint and produces edge-ready binaries with documented quality bounds. Lineage tracked from source weights to deployed binary.

Multi-modal applications

Vision + audio + time-series + text + telemetry, all addressable through the same SDK. Compose multi-model agents on-device — Qwen3-VL for vision, PANNs for acoustic, TimesFM for time-series — and let the runtime route inference to the right silicon.

Federated fine-tuning

When user data shouldn't leave the device, ExecuTorch + NVIDIA FLARE handle on-device fine-tuning and aggregate model deltas centrally. The pattern Meta + NVIDIA shipped in 2025 — adopt it via Furcate's SDK without reimplementing the protocol.

Wasm serverless edge

Wrap your application logic as WASI 0.3 components. Cold-start is 1-5 ms vs 100ms-1s+ for containers. Cloudflare-class scale (10M+ Wasm requests/sec across 300+ edge locations) is the proof-point; Furcate brings that pattern to your fleet.

Local development → production parity

The same SDK runs on your laptop, on a Pi on your desk, and on the customer's air-gapped industrial gateway. Same APIs, same observability, same provenance trail. No surprises in production.