Intel Lava Is Archived: What to Use Instead
Intel's Lava neuromorphic framework, the closest thing Loihi 2 development has had to an official SDK, now sits archived on GitHub with no public successor. Here is what that actually means, and what still works if you were building on it.
If you maintain a project that depends on Intel's Lava framework for Loihi 2 development, you have probably already noticed: the lava-nc GitHub organization's repositories, including the core Lava framework and Lava-DL, are now marked archived. Intel has signaled that a new SDK is coming alongside next-generation Loihi hardware, but nothing has shipped publicly yet. For now, Lava is in a frozen, read-only state - the code still runs, but no fixes, no new features, no responses to open issues.
What "archived" actually means for Lava
An archived GitHub repository is not deleted. You can still clone lava-nc/lava and lava-nc/lava-dl, install them, and run existing projects. What changes is everything forward-looking: no merged pull requests, no triaged issues, no compatibility updates as PyTorch, Python, or CUDA versions move forward. If you hit a bug today, you are debugging it yourself with no upstream help coming.
This matters more for Lava specifically because of how central it was to the Loihi 2 workflow. The typical path - train a PyTorch-style model with Lava-DL's SLAYER module, export to HDF5, convert to Lava Processes via NetX, then run on Loihi2SimCfg or Loihi2HwCfg - depended on Lava as the only supported bridge between PyTorch and the chip. With that bridge archived, every step downstream of it is also effectively frozen.
Why this happened (and why it's not just Lava)
Loihi 1 and Loihi 2 were always research systems, not commercial products - access required joining the Intel Neuromorphic Research Community (INRC), and the open-source core only ever supported CPU simulation by default, with the actual hardware backend gated behind that membership. Lava's archival is consistent with Intel repositioning around next-generation Loihi hardware and tooling, but until that tooling ships, anyone without an active INRC relationship is left without an official path forward.
This is the exact problem that motivated NeuroCUDA, an open-source PyTorch-to-spiking-neural-network compiler: neuromorphic software has tended to be vendor-controlled, gated behind research-community membership, and prone to exactly this kind of disruption when a single vendor's roadmap shifts.
What still works if you were on Lava
- Existing Lava-based code keeps running. Archived does not mean broken. If your pipeline is already trained and deployed, nothing changes immediately.
- PyTorch-native SNN libraries are unaffected. SpikingJelly, snnTorch, and Norse are independent of Lava and continue to be actively maintained.
- NIR export gives you a vendor-neutral fallback. Models exported to NIR (Neuromorphic Intermediate Representation) are not locked to Lava's format, and can move to other NIR-supporting simulators as those mature.
- Simulator-validated Loihi 2 deployment is possible without Lava. NeuroCUDA ships a Loihi 2 IF-neuron simulator backend validated against Intel's own published neuron equations rather than wrapping the Lava SDK, so you can develop and validate neuron-level behavior without an active Lava dependency or INRC access.
Where NeuroCUDA fits as a Lava alternative
NeuroCUDA does not claim to replace Lava on physical Loihi 2 silicon - that distinction matters and the project is explicit about it. What it does provide: a pip install neurocuda path that takes a trained PyTorch model, converts it to a spiking neural network using QCFS calibration and BPTT fine-tuning, and runs it on GPU, CPU, or the Loihi 2 simulator, with NIR export so the resulting model isn't tied to any single tool going forward.
The numbers are measured, not projected: a 3-layer CNN converted for N-MNIST reaches 99.88% ± 0.02% accuracy against a 99.70% ANN baseline, and a ResNet-18 on CIFAR-10 converts with a 0.95% accuracy gap. The Loihi 2 simulator backend is checked against Intel's published Loihi neuron equations across 100,000+ comparisons with zero deviations - a spec-conformance check, not a claim of running on Intel's Lava SDK or physical silicon.
| Need | Lava (archived) | NeuroCUDA |
|---|---|---|
| Install path | conda + INRC membership for hardware access | pip install neurocuda |
| Active maintenance | Archived, no updates | Active |
| PyTorch model in, SNN out | Via SLAYER + HDF5 export | One convert() call |
| Loihi 2 backend | Lava SDK + physical hardware (INRC-gated) | Simulator vs. published neuron equations |
| Vendor-neutral export | Lava-specific format | NIR export, bit-exact on residual graphs |
The honest limits, on both sides
Neither side of this comparison is complete. Lava's archival doesn't erase the years of work that went into Lava-DL's SLAYER training pipeline or its NetX deployment path - if and when Intel ships a successor SDK, Lava-trained models and workflows will likely have a migration path. And NeuroCUDA's Loihi 2 backend is, today, a simulator validated against published equations, not a substitute for running on physical Loihi 2 silicon through an INRC hardware allocation. If your work requires physical Loihi 2 hardware specifically, NeuroCUDA's simulator gets you accurate neuron-level behavior for development and testing, not a hardware deployment.
What NeuroCUDA does remove is the dependency on a single vendor's SDK staying maintained. A PyTorch model converted and exported to NIR today is not stranded if any one tool in the chain goes archived tomorrow.
Sources & further reading
lava-nc/lavaandlava-nc/lava-dlGitHub repositories, archived status observed June 2026- Intel Neuromorphic Research Community (INRC) membership requirements for Loihi 2 hardware access
- NeuroCUDA source and verified benchmark results, github.com/Krishnav1/neurocuda
- NIR (Neuromorphic Intermediate Representation) specification, arXiv:2311.14641
Frequently asked questions
Is Intel Lava archived?
Yes. The lava-nc GitHub organization's repositories, including the core Lava framework and Lava-DL, are archived as of 2026, with Intel signaling that a new SDK for next-generation Loihi hardware is coming but not yet released.
What should I use instead of Lava for Loihi 2 development?
There is no direct drop-in replacement that targets physical Loihi 2 silicon outside Intel's own tooling. For PyTorch-based development and simulation, options include NIR-compatible libraries like snnTorch, Norse, and Rockpool, and compilers like NeuroCUDA that convert PyTorch models to spiking networks and validate against Loihi 2's published neuron equations in simulation.
Can I still access archived Lava repositories?
Yes, archived GitHub repositories remain readable and clonable, but they no longer receive updates, bug fixes, or issue responses from maintainers. Existing Lava-based projects will continue to run but will not benefit from active development.
Does NeuroCUDA require Lava or INRC membership?
No. NeuroCUDA is pip-installable and does not require Intel Neuromorphic Research Community (INRC) membership. Its Loihi 2 backend is a simulator validated against Intel's published Loihi neuron equations, not a wrapper around the Lava SDK.