Voltaic has reached the third milestone on its roadmap: validated, live and scaling. The architecture is locked. From here, the work is about scale, production hardening and, soon, giving the brain a body. Here's where we've been, where we are and where we're going.

1. Shipped: the spiking core

The first milestone was the hardest to believe in: a from-scratch spiking language model, trained on real language data and proven to learn. Not a conventional model with spiking added on, but binary spikes, leaky integrate-and-fire neurons and STDP-weighted attention all the way through. Plenty of people told us spiking networks couldn't do language. Getting one to learn real language was the proof that the rest of the roadmap was possible.

2. Shipped: memory that holds and sharpens

A model that learns continuously is only useful if it doesn't forget. The second milestone delivered three pieces that work together:

  • HSC v2 Living Memory, which tracks the importance, velocity and clarity of every synapse and hardens the ones that matter.
  • MemoryForge, which turns consolidation into a working memory: keeping what's important, merging what's related and letting noise fade.
  • HULL, the Accumulative Hull Neuron Model, which keeps a stability hull around mastered skills.

With them, forgetting on CORe50-NI dropped from 1.20% to 0.04%, while accuracy rose. See the benchmarks.

3. We are here: validated, live and scaling

Voltaic has been evaluated on CLBench v10 and is running live as a multimodal app at voltaic.live. Today it supports:

  • Chat with a model that remembers what you taught it.
  • Teaching facts and skills at inference time, with no retraining run.
  • Voice input into the same spiking brain.
  • Camera-gesture learning: show it a gesture and it learns it, in the same weights as everything else.
  • External model orchestration: the spiking core can route to an open model on your own hardware or a frontier model in the cloud, and fuse the answer with its own memory.

"Architecture locked" means the fundamental design won't change. What changes now is scale, from 355M up to 7B parameters on a single GPU, and the engineering that production deployments need: reliability, monitoring, deployment options and the governance that comes from running underneath Nexus and Halo.

4. Next: embodied intelligence

Most robots have brains that are frozen. Through Verkko Robotics, Voltaic gets a body: robots that learn skills on the job. Paired with HULL, a robot keeps a stability hull around the skills it has mastered, so it can pick up something new without unlearning how to walk. Robots will be taught by demonstration, with hands, voice and vision in one brain, and will learn in the field rather than in a retraining run. This is what powers Verkko Kinetic.

5. Horizon: our own hardware

Spiking neurons only fire when there's something new to process. Our horizon is to deploy Voltaic on Verkko's own proprietary hardware, designed around the spiking core, bringing continual learning to the edge at far less power, privately and on a battery.

Why the order matters

Each milestone depends on the one before. There's no point embodying a model that forgets, and no point scaling one that can't prove what it knows. Locking the architecture now means everything we build next, in Nexus, Halo and Kinetic, stands on a foundation that won't shift underneath it.

Read more on the Voltaic page or try it at voltaic.live.