Next Brain Research
Build the evidence
before the claim.
We use the execution platform to study which neural details must remain explicit, which can be approximated, and how results survive changes in runtime and scale.
01 · ModelEquations and structure
02 · InterventionChange one boundary
03 · EvidenceRecord result and limits
Featured publication · Preprint v2
A Unified Intermediate Representation and Execution Architecture for Heterogeneous and Distributed Neural Simulation
A shared model-to-execution architecture, with measurements across CPU, GPU, browser and distributed systems. Xinjun Li · Next Brain · 10 October 2026.
Research map
One stack.
Four connected questions.
We organise the work by scientific question—not by repository, backend or demo.
Current foundation
Execute faithfully
How can researcher-authored equations become inspectable execution plans with reproducible results across runtimes?
Current frontier
Scale responsibly
How far can models travel across local GPUs, browsers and clusters without losing identity, provenance or resource bounds?
Active research
Compress what we can
Which local dynamics can proxy states and procedural connectivity replace—and which motifs must remain explicit?
Exploratory
Explain what emerges
Can perception, adaptation and safety behaviour be isolated through interventions that separate capability from mechanism?
Evidence ledger
What we know.
What remains open.
Every public claim should carry its model identity, code version, environment, result and boundary.
ProgrammeCurrent evidenceBoundary / next test
Engineering verifiedRuntime substrate
A shared model identity reaches native Rust, GPU, browser and MPI execution paths.
Broader model compatibility and independent replication remain open.
Boundary explicitMulti-area model
Full-scale 4.13M-neuron / 24.1B-synapse bounded engineering runs have completed.
Scientific equivalence, repeatability and speed / cost advantage are not yet closed.
Scientific evidenceFlyWire vision
A fixed real-connectome network carries task information readable by an external classifier.
Matched random rewiring, multiple full seeds and internal plasticity are still needed.
Boundary foundApproximate dynamics
Richer proxies and selective retention expose a measurable fidelity–cost frontier.
The E1 plateau and 393 Hz artefact motivate stateful proxies and automatic motif selection.
Long-term blueprint
The roadmap is a sequence of testable questions—not a promise of inevitable scale.
- 01Verified execution substrateModel identity · portable plans · bounded runtimes
- 02Large-scale explicit modelsIndependent references · multiple seeds · cost accounting
- 03Selective structureProcedural connectivity · motif retention · fidelity frontier
- 04Stateful local proxiesShort memory · adaptive dynamics · boundary recovery
- 05Cross-scale memoryPersistent state · route-based reactivation · continual adaptation
Research principleEngineering results and scientific interpretation stay separate.
A faster runtime is an engineering result. A model preserving the same scientific meaning is a separate claim—and needs its own evidence.
Follow the work