Skip to content

Overview

Bitweaver is a platform by DeepGate that compiles PyTorch models to optimized C libraries for microcontrollers and profiles them on real hardware.

You give Bitweaver a model in the form of a schema JSON. You get back a deployable C library and measured numbers from the exact MCU you're targeting: inference time, a per-layer breakdown of inference time, memory footprint, and flash.

Bitweaver also hosts a Model Zoo of pre-trained models, which can be directly integrated into your embedded application. Alternatively, pull a model with the DeepGate SDK, fine-tune it for your task, and send it through the same compile-and-benchmark pipeline as a model you trained from scratch.

Who it's for

Engineers and teams shipping AI to microcontrollers who need:

  • A robust path from PyTorch to a highly optimized, MCU-ready C library without hand-porting layers.
  • Real, measured numbers from the target hardware, not host-side estimates.
  • A shared place to compare model variants across boards and iterations.

What it isn't

  • Not a training platform: Bring a trained model, or a pre-trained / fine-tuned model from the Model Zoo.
  • Not a runtime: Bitweaver produces a static library that you link into your embedded application.
  • Hardware scope: See Supported Hardware for the current list of tested MCUs.

Workflow

Bitweaver workflow

The workflow is iterative:

  • If the model misses your defined requirements (e.g. accuracy), stay in the developer environment and retrain or fine-tune.
  • If the model's on-MCU performance (inference time, memory, flash) misses your requirements, go back and redesign.

Bitweaver speeds up this cycle by automating optimization, profiling, and deployment.

Benchmark lifecycle

Every upload creates a benchmark job that moves through these states:

Stage What's happening
build_queued Waiting for a free buildworker.
building Buildworker is optimizing and compiling the model.
run_queued Compile succeeded; waiting for target hardware to free up.
running Model is executing on the MCU and being profiled.
completed Finished successfully. Metrics and artifacts are available.
build_failed / run_failed Stopped with an error. See the Submission view for details.
cancelled You cancelled the job before it finished.

Managing jobs

Jobs are managed from the Run History panel on the project dashboard.

  • Cancel a job at any point before it reaches a terminal state.
  • Delete a single job, or an entire Submission (a group of jobs uploaded together).
  • Failures surface inline in the Submission view with the stage and error message.

Core concepts

Project: A named container for related benchmarks. One project, one dashboard, one shared set of default metrics and target boards. Use projects to compare variants of the same model, or to isolate work across teams.

  • Dashboard: Per-project view with a Performance Map (metrics across runs) and Run History (every job, its state, and its artifacts).

  • Project Settings: Owners can edit the name, description, visibility, default metrics, and default benchmark targets.

Model Zoo: Pre-built models you can deploy directly onto your MCU. Each Zoo model can be pulled and fine-tuned with the DeepGate SDK.

Schema JSON: The file the SDK writes alongside your model weights. It describes the model's layout so Bitweaver can compile and benchmark it. See Quick Start for where it lands.