SpikeForge is my Python toolkit for building and testing spiking neural networks. It stopped behaving like one big project during this release cycle: the core toolkit, target adapters, model hub and SpikeForge Dashboard all moved forward as separate packages and applications.
The core project released 0.4.0 and 0.5.0, then kept publishing the smaller package releases that make the ecosystem usable. The event path got a real train/test split. Quantization work moved into the drift checks. Shared logging came from capsize-commons, and the CI path gained a fast leaf lane for the parts that do not need the whole workspace.
The package layout is now a practical map of the work:
spikeforge holds the training and experiment workflow.spikeforge-targets holds target-specific conversions.spikeforge-hub owns the offline-first model catalogue.spikeforge-dashboard contains the web dashboard and desktop application.
That separation is useful, not just a tidy repository list. Someone experimenting in Python does not need to install the desktop app. Someone browsing models does not need the server. Each piece can ship when it is ready, with compatibility recorded instead of implied.
The stack is Python, PyTorch and snnTorch around the core, with a web dashboard and desktop application around it. The model hub is deliberately cache-friendly, so a catalogue can be inspected and used without turning every experiment into a network operation. SpikeForge Dashboard gives the same work a visual entry point and provides a place to inspect the experiment rather than only its final number.
The project page has the current package links and release notes: spikeforge.net. The desktop releases are collected on itch.io, and the source is on GitHub.
Install the core package, choose one of the examples, and keep the first run small. Start with one dataset and a short training window. Then move the same experiment through the target adapter and save the result in the hub catalogue. The point of the exercise is not a leaderboard number. It is seeing the path from an experiment to a named, repeatable artefact.
A follow-up article walks through a small event-driven classifier. The four-week overview covers the rest of the Capsize work from the same period.