Spiking network software in 2026
The open-source software for spiking neural networks in 2026: what each tool is for, its licence, and how a model reaches a chip
A neuromorphic chip is only as useful as the software that trains and loads it. We read the READMEs, licence files and release records of eleven open-source projects: which ones train networks, which ones simulate brains, which ones target hardware, how models move between them, and the licence terms that matter if you build a product.
Why the software matters
A spiking neural network is trained on an ordinary computer and then run somewhere else: on a CPU, a GPU or a neuromorphic chip. The tools in between decide what you can build. They decide which neuron models you can use, how you train them, whether the result can be loaded onto a chip, and on what licence terms you can ship it.
This piece is a map of those tools as they stand in November 2026. Every fact comes from a project's own README, licence file or release record on PyPI, read on 22 November 2026. The descriptions are the projects' own words, summarised. We have not benchmarked these libraries against each other, so there are no speed or energy figures here.
Three kinds of tool
Training libraries treat spiking neurons as layers in a deep learning framework, so you can train a network with gradient descent on a GPU. snnTorch, Norse, SpikingJelly and Sinabs build on PyTorch. Spyx builds on JAX and Flax NNX. Rockpool offers layers backed by PyTorch, JAX or Brian.
Simulators come from neuroscience. They model how networks of neurons behave over time, rather than training them for a task. Brian 2 describes itself as a clock-driven simulator written in Python. NEST focuses on the dynamics, size and structure of neural systems. NEST's README says it runs from a laptop to the largest supercomputers and was first released in 1994 (as SYNOD).
Hardware toolchains turn a trained network into something a chip can run. Intel's Lava and Lava-DL targeted the Loihi chips. Two SynSense projects target the company's own chips: Rockpool for Xylo, and Sinabs for Speck and DYNAP-CNN.
| Project | What it is for | Built on | Licence | Latest release |
|---|---|---|---|---|
| snnTorch | Training spiking networks with gradients | PyTorch | MIT | 1.0.0, June 2026 |
| Norse | Deep learning with spiking networks | PyTorch | LGPL-3.0 | 1.1.0, March 2024 |
| SpikingJelly | Large-scale training and inference, conversion from ordinary networks | PyTorch | Apache-2.0 | 2.0.0, October 2026 |
| Spyx | Compact training library | JAX, Flax NNX | MIT | 1.0.0, July 2026 |
| Sinabs | Spiking convolutional networks, deployment to DYNAP-CNN and Speck chips | PyTorch | Apache-2.0 | 3.1.3, February 2026 |
| Rockpool | Signal processing with spiking networks, deployment to Xylo chips | PyTorch, JAX or Brian | AGPL-3.0 or later | 3.1.0, July 2026 |
| Lava | Building applications for neuromorphic hardware (archived) | Python | BSD-3-Clause and LGPL-2.1 or later | 0.10.0, August 2024 |
| Lava-DL | Training and deploying event-based deep networks (archived) | PyTorch | BSD-3-Clause | 0.6.0, August 2024 |
| Brian 2 | Simulating spiking networks | Python | CeCILL-2.1 | 2.11.0, October 2026 |
| NEST | Simulating large networks of spiking neurons | Python interface or stand-alone program | GPL-2.0 or later | 3.10.0, June 2026 |
| NIR | A shared format for moving models between tools | Python | BSD-3-Clause | 1.0.8, July 2026 |
How the training libraries train
A spike is all or nothing, so it has no useful slope for gradient descent. The training libraries work around this in two ways, and the Lava-DL README states the trade-off plainly.
Direct training uses the precise timing of spikes. Lava-DL calls it very accurate and efficient, but slow and costly to train. In practice this means surrogate gradients: during training, the spike's missing slope is replaced with a smooth stand-in. snnTorch ships these stand-ins as a module of their own, so you can swap one for another.
Conversion trains an ordinary network first, then swaps its activations for spiking neurons whose firing rates stand in for the values. It reuses fast, familiar training. The cost, in Lava-DL's words, is that converted networks typically need more time steps to reach an answer than directly trained ones. SpikingJelly lists conversion among its core features. Our own spiking conversion study found the same pattern in simulation: accuracy improved only as the network ran for more time steps.
Some tools add other routes. Spyx offers gradient-free training by neuroevolution alongside surrogate gradients. Lava-DL lists a plug-in for local learning, where each layer learns from its own signal instead of a gradient passed back through the whole network.
A detail that matters on small machines: snnTorch says its neuron models are written as recursive functions, so it does not need to store every neuron's membrane potential over time to compute a gradient. Its README says this keeps networks small enough to train on a CPU where needed.
Moving a model from one tool to another
For years, a network trained in one library could not easily run in another or on a chip from another maker. The Neuromorphic Intermediate Representation (NIR) is a shared set of building blocks that tools can write to and read from. It was conceived at the 2023 Telluride Neuromorphic Workshop and described in a 2024 paper in Nature Communications. Its README says it is supported by 9 simulators and 5 hardware platforms.
Support is not symmetrical. The table below is the README's own list. Lava-DL and SpiNNaker2 read NIR but do not write it, so they are destinations, not sources.
| Tool | Writes NIR | Reads NIR | Hardware it reaches |
|---|---|---|---|
| snnTorch | Yes | Yes | None of its own |
| Norse | Yes | Yes | None of its own |
| Spyx | Yes | Yes | None of its own |
| Nengo | Yes | Yes | None listed |
| Rockpool | Yes | Yes | SynSense Xylo |
| Sinabs | Yes | Yes | SynSense Speck |
| hxtorch, jaxsnn | Yes | Yes | BrainScaleS-2 |
| Lava-DL | No | Yes | Intel Loihi |
| SpiNNaker2 | No | Yes | SpiNNaker2 |
SpikingJelly is not in that list, but its own README links tutorials for NIR exchange, Lava exchange and deployment to Lynxi chips. A shared format is a real step, but it describes the network, not how a given chip will run it. Chips can differ in the neuron types and number formats they support, so a moved model still has to be checked on the target.
The licence terms that matter for a product
If you only do research, most of these licences will not get in your way. If you ship a product, they differ a lot. This is a summary for engineers, not legal advice.
Permissive licences: snnTorch and Spyx (MIT), SpikingJelly and Sinabs (Apache-2.0), Lava-DL and NIR (BSD-3-Clause). SpikingJelly's licence guide sums up what Apache-2.0 allows: research, commercial use, change and redistribution, on the licence's terms.
Copyleft licences, which attach conditions to sharing changed or combined code: Norse is LGPL-3.0, NEST is GPL-2.0 or later, Brian 2 uses CeCILL-2.1, and Rockpool is AGPL-3.0 or later. Rockpool's README says commercial licences are available on request.
Mixed: the Lava licence file puts the core, processes, utilities and tutorials under BSD-3-Clause, and the compiler and runtime under LGPL-2.1 or later. Its README calls the framework permissively licensed, so read the licence file, not only the summary.
Read the full terms, or ask a lawyer, before a copyleft library ends up inside something you sell.
SpikingJelly moved to Apache-2.0 for its version 2. Its licence guide says that older tags and published releases keep their original licensing records, so check which version you install.
Maintenance is a risk to plan for
Intel's Lava and Lava-DL repositories are archived. The READMEs say Intel will not provide maintenance, bug fixes or new releases, and that Intel is developing a new SDK for its next-generation Loihi chip, built on open-standard AI frameworks. Code written for Lava today has no maintained path forward from Intel.
Release records tell a mixed story elsewhere. SpikingJelly, Brian 2, Rockpool, snnTorch, Spyx, NIR and NEST all published releases in 2026. Norse's latest release on PyPI is from March 2024; its repository may hold newer code, but there is no newer packaged release.
What we use, and what we have not done
Our simulation studies use none of these libraries. They are written in plain JavaScript with no dependencies and fixed random seeds, so anyone can re-run them and get the same numbers. That is right for small studies and wrong for anything larger.
We have not benchmarked these tools against each other, trained the same network in each, or run any of them on neuromorphic hardware.
Sources
- snnTorch README (project README)
- snnTorch MIT licence file
- snnTorch release record on PyPI
- Norse README (project README)
- Norse LGPL-3.0 licence file
- Norse release record on PyPI
- SpikingJelly README and licence guide (project README)
- SpikingJelly Apache-2.0 licence file
- SpikingJelly release record on PyPI
- Spyx README (project README)
- Spyx MIT licence file
- Spyx release record on PyPI
- Sinabs README (vendor README, SynSense)
- Sinabs Apache-2.0 licence file
- Sinabs release record on PyPI
- Rockpool README (vendor README, SynSense)
- Rockpool AGPL-3.0 licence file
- Rockpool release record and licence classifier on PyPI
- Lava README and archive notice (vendor README, Intel)
- Lava licence file (BSD-3-Clause and LGPL-2.1 or later)
- Lava-DL README and archive notice (vendor README, Intel)
- Lava-DL BSD-3-Clause licence file
- Lava release record on PyPI
- Lava-DL release record on PyPI
- Brian 2 README (project README)
- Brian 2 CeCILL-2.1 licence file
- Brian 2 release record on PyPI
- NEST README (project README)
- NEST GPL-2.0 or later licence file
- NEST release record on PyPI
- NIR README and support table (project README)
- NIR BSD-3-Clause licence file
- NIR release record on PyPI