Voltaic is a from-scratch spiking neural network, built from binary spikes, LIF neurons and STDP attention. It absorbs new knowledge at inference time and forgets far less than conventional models. Old skills often sharpen as it learns new ones. Now we're giving it a body.
Down from 1.20% for a standard model, while accuracy rose too.
On Split ImageNet-1k, with no replay buffer.
Runs on a single GPU today.
Earlier tasks are never re-shown during training.
The model you talk to today is the same one that was shipped months ago. It can't learn from you. Teaching it something new means a full retraining run, and that run tends to overwrite what it already knew. The industry calls this catastrophic forgetting.
Voltaic takes the opposite path: a brain-inspired network that learns continuously, protects what matters, and can show you exactly which knowledge is safe.
Teach facts, skills and gestures at inference time. There's no retraining run and no redeploy; knowledge lands while the model is running.
Text, voice and camera flow into the same weights. It's a single spiking representation for every sense, not three APIs stitched together.
Hierarchical Synaptic Consolidation hardens important synapses. We can show you the forgetting curve, which knowledge is protected, and why.
Every input becomes spikes, flows through the spiking transformer, and is woven into a living memory that grows with use.
Text, image, audio and structured data become binary spike tensors over simulated neural time.
LIF and Adaptive-LIF neurons fire across 16 timesteps, with STDP-weighted attention.
HSC v2 tracks each synapse's importance, velocity and clarity, so persistence has to be earned.
Assembles the relevant memory. The model answers, explains itself and keeps learning.
Hardens the synapses behind skills you rely on and leaves the rest free to adapt. Read more →
Keeps a geometric "stability hull" around mastered skills, so new learning doesn't unlearn old abilities.
Keeps what's important, merges what's related and lets noise fade. Visible, searchable and deletable.
On public continual-learning benchmarks, with no replay buffer, Voltaic retains 2.6× to 4.2× more knowledge than dense 7B baselines (Qwen and Mistral) as new domains arrive.
4.2× the dense baselines. Retained is accuracy on earlier tasks after every task in the sequence has been learned, so higher is better and 100% would mean nothing was forgotten.
That's 2.6× more knowledge retained.
HULL also lifted average accuracy from 95.01% to 95.23% at the same time.
CLBench v10, no-replay sequential protocol, single A100, 30 June 2026. Vision tests: 8× A100, 14 June 2026. For the vision benchmarks, images are encoded with frozen, pretrained DINOv2 features; continual learning happens in Voltaic's spiking core on top of them. Single seed (seed 0); multi-seed results to follow.
No, not in the way the rest of the industry means it. The spiking core adapts its own weights locally, inside your workspace, and only from what you deliberately teach it.
| "Training the model" in frozen AI | Voltaic continual learning | |
|---|---|---|
| When | Offline, in a distant data centre | Live, as you use it, scoped to your workspace |
| Where your data lives | Pooled into a shared dataset | Stays in your workspace, unless you allow cloud routing |
| What learns | A single model shipped to everyone | Only your workspace's memory and weights |
| Who benefits | The vendor's next model | You, and only you |
| Opting out | Buried in settings, if it exists at all | Nothing to opt out of: privacy is the default |
Voltaic's spiking core doesn't have to answer alone. It can orchestrate an open model on your own hardware or a frontier model in the cloud, and fuse their output with its own living memory.
A from-scratch spiking language model, trained on real language data and proven to learn.
HSC v2 Living Memory, MemoryForge and HULL. Forgetting drops to 0.04%.
Proven on CLBench v10 and running live as a multimodal app. The architecture is locked.
Verkko Robotics gives the brain a body: robots that learn skills on the job.
Voltaic deployed on Verkko's proprietary hardware, designed around the spiking core for far lower power use. We will publish measured energy figures once we have them.
Verkko Robotics is building robots with Voltaic. 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. This is what powers Verkko Kinetic.
It learns new skills live, from any sense, without forgetting, and it can explain itself. It runs on a single GPU.