TALON (Tactical AI at Low-power On-device Nodes) is a neuromorphic computing SDK for Type 1 Compute hardware. It connects machine-learning frameworks to neuromorphic deployment, for spiking neural networks and hybrid architectures alike. I built three of its pieces: the simulation engine, the IR visualiser and the binary encoder. The encoder stores a ten-layer network in about 200 bytes instead of about 200 KB.

~1000×Smaller: a ten-layer network goes from 200 KB to 200 B
7Packages in the SDK
36IR primitives across SNNs, ANNs and hybrids
2My fixes merged upstream into NIR

01Overview

TALON mirrors the ONNX approach. A PyTorch model is exported to a TALON IR graph with bridge.to_ir(), then written to a portable .t1c file with ir.write(). From there, separate packages map the graph onto hardware, compile, simulate and profile it, or visualise it. The split lets ML engineers export models without knowing the hardware, and lets hardware teams consume a standard IR without touching PyTorch.

The TALON pipeline, after the project's documentation.

02Packages

PackagePyPIPurpose
talon.sdkt1c-talonMeta-package and the talon / t1c command-line tools
talon.irtalon-irCore IR primitives and HDF5 serialisation
talon.bridgetalon-bridgePyTorch export and import
talon.viztalon-vizInteractive graph and spike visualisation
talon.graphtalon-graphPartitioning, placement, routing and resource allocation
talon.backendtalon-backendCompilation, CPU simulation and the hls4ml FPGA backend
talon.iotalon-ioEvent streaming, format conversion and neural encoding

The IR covers spiking, conventional and hybrid networks, with standard activations and normalisation layers. Graph algorithms run on RustworkX, the CPU simulator estimates latency and energy, and the event I/O package decodes EVT2 and EVT3 streams at more than 5 million events per second.

03My work

Simulation engine
A lightweight execution engine with optimised primitives for spike-based computation, handling temporal dynamics and event-driven processing with minimal memory overhead.
IR visualiser
Interactive tools for debugging and analysing spiking networks, including real-time spike-pattern visualisation and performance profiling.
Binary encoder
A compact serialisation format for efficient inference. It stores a ten-layer network in about 200 bytes instead of about 200 KB.

04Upstream contributions

TALON's neuron dynamics follow the Neuromorphic Intermediate Representation (NIR), and I've contributed fixes to NIR's reference implementation. Two were merged in March 2026. The first repaired a type-check guard in read() that tested for an attribute on a dictionary and so could never fire. The second added structural validation that names dangling and duplicate edges before type inference begins. A third, still in review, resolves the types of scalar (zero-dimensional) neuron parameters from graph context and stops HDF5 from trying to compress zero-dimensional datasets.

05References

  1. TALON Ecosystem: Introduction. Type 1 Compute documentation, retrieved 9 October 2026.
  2. Faras Siddiqui. Curriculum vitae (PDF).
  3. Faras Siddiqui. Projects. Earlier version of this site.
  4. Faras Siddiqui. fix: use dict key lookup instead of hasattr in read(). neuromorphs/NIR pull request #187, 25 February 2026.
  5. Faras Siddiqui. fix: add structural graph validation before type inference. neuromorphs/NIR pull request #188, 25 February 2026.
  6. Faras Siddiqui. fix: resolve scalar parameter types from graph context during inference. neuromorphs/NIR pull request #189, 25 February 2026.