BrainX EcosystemBuild, simulate and train brain models at every scale

BrainX Ecosystem brings together open-source JAX libraries for building brain models at every scale, fitting them to data with gradients, and running them on CPU, GPU or TPU.

Install Browse packages

One runtime from ion channel to whole brain

The same state objects, integrators and compiler run every model, so a detailed neuron can sit inside a network that sits inside a whole-brain model.

What you can do with it

Full-scale simulation
Model ion channels, dendrites, spiking networks and whole-brain dynamics with one set of tools.
AI compatibility
Use Flax, Optax and Orbax inside a brain model. It is all JAX, so they work together directly.
Differentiable optimization
Fit models to recordings or train them on tasks with backprop through time, surrogate gradients or online learning.
Powered by JAX
Compile models with jit and run the same code on a laptop CPU, a GPU or a TPU.

Eight packages, three layers

Install only what your model needs. Every package builds on BrainState.

ModelingPick by scale

BrainSpikeSpiking networks of point neurons
BrainCellBiophysically detailed neurons
BrainMassWhole-brain neural-mass models

Shared infrastructureUsed by every model

BrainUnitPhysical units for JAX
BrainEventEvent-driven sparse operators
BrainTraceOnline gradient learning
BrainToolsVisualization and utilities

RuntimeUnder everything

BrainStateState management and compilation on JAX

Using an AI coding agent? The BrainX skill teaches it the ecosystem. See every package

Why pick BrainX

How BrainX compares with other brain simulators. Partial means the feature exists through an add-on or with limits; the notes say which.

Capability	BrainX	NEURON	Brian2	NEST	NetPyNE	TVB
Multi-compartment neurons	Yes	Yes	Yes	Partial^{note 1}	Yes	No
Large spiking networks of point neurons	Yes	Yes	Yes	Yes	Yes	Partial^{note 2}
Whole-brain neural-mass models	Yes	No	No	Partial^{note 3}	No	Yes
Physical-unit checking	Yes	Partial^{note 4}	Yes	Partial^{note 5}	Partial^{note 4}	No
GPU or TPU	Yes	Partial^{note 6}	Partial^{note 7}	Partial^{note 8}	Partial^{note 6}	Partial^{note 9}
Multi-node (MPI) simulation	Yes	Yes	No	Yes	Yes	Partial^{note 10}
Differentiable end to end	Yes	No	No	No	No	Partial^{note 11}
Built-in online gradient learning	Yes	No	No	Yes	No	No
Works with Flax and Optax	Yes	No	No	No	No	Partial^{note 11}

NEST: one compartmental model (cm_default).
TVB: by co-simulating with NEST, ANNarchy or NetPyNE through tvb-multiscale.
NEST: rate neurons, but no whole-brain or connectome tooling.
NEURON and NetPyNE: modlunit checks units in NMODL mechanism files only.
NEST: in models written in NESTML only.
NEURON and NetPyNE: GPU through CoreNEURON, with some mechanism restrictions.
Brian2: GPU through the separate Brian2CUDA or Brian2GeNN packages.
NEST: GPU through the separate NEST GPU project.
TVB: RateML generates CUDA kernels for parameter sweeps.
TVB: MPI spreads independent parameter-sweep runs, not one simulation.
TVB: through the separate JAX library vbjax.