Neuromorphic intelligence · live

A mind that learns while you use it.

Voltaic is a from-scratch spiking neural network — binary spikes, LIF neurons, STDP attention. It absorbs new knowledge at inference time and barely forgets — old skills often sharpen as it learns new ones. Now we're giving it a body.

Where we're going → See how it works
0 spiking parameters
0.71× avg drift · old skills sharpen
∞ continual memory

Today's AI is frozen at training time.

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.

What Voltaic is

Intelligence that compounds.

Instead of dense, continuous activations, Voltaic processes information as discrete spike events — the same sparse, event-driven dynamics your neurons use. Three things fall out of that design.

⚡

Learn live

Teach facts, skills, and gestures at inference time. No retraining run, no redeploy — knowledge lands while the model is running.

🧠

One brain

Text, voice, and camera flow into the same weights. Not three APIs stitched together — a single spiking representation for every sense.

🛡️

Prove it

Hierarchical Synaptic Consolidation hardens important synapses. We can show you the forgetting curve, which knowledge is protected, and why.

AI sovereignty

Your intelligence. Your data. Your hardware.

The world's most capable AI lives behind a handful of foreign APIs — frozen models you rent, fed by data you hand over, on infrastructure you'll never control. That's not a strategy; it's a dependency. Voltaic is built so intelligence stays yours.

01

Runs on your own metal

A 355M–7B spiking brain runs on a single GPU today and targets neuromorphic silicon at the edge tomorrow. No rented black box, no per-token tax, no kill switch in someone else's data center.

02

Your data never leaves

Voltaic learns locally, at inference time. Knowledge compounds on-premise — nothing is phoned home to train a competitor's model. Sovereignty over your data is the default, not a paid tier.

03

Built from scratch

Not a thin wrapper over a frozen foundation model. The architecture, the weights, the Living Memory — all ours, all auditable. You own the whole stack down to the spike.

04

Sovereign by design

Engineered at the Verkko Robotics Research Lab in Europe, for organizations and nations that refuse to outsource their cognition. Independence, baked into the architecture.

“A nation that rents its intelligence doesn't own its future. We build minds you can keep.”
How it works

From any sense to a single spiking mind.

Every input becomes spikes, flows through the spiking transformer, and is woven into a living memory that grows with use.

01

Encode

Text, image, audio & structured data → binary spike tensors over simulated neural time.

02

Spike

LIF / Adaptive-LIF neurons fire across 16 timesteps with STDP-weighted attention.

03

Consolidate

Living Memory tracks each synapse's importance, velocity & clarity — persistence is earned.

04

Respond

ContextFlow assembles relevant memory; the model answers — and explains itself.

# the whole loop, in one line of intent
raw_input  →  spike_encode()  →  spiking_transformer()  →  memory_forge()  →  answer
# …and the model keeps learning from every interaction, without forgetting.
Orchestration

One core. Every model at its command.

Voltaic's spiking core doesn't have to answer alone. It can orchestrate external models — an open model on your own hardware or a frontier model in the cloud — and fuse their output with its own living memory. You choose how much control to hand over: let the core decide autonomously, or set your own routing rules. Either way, the spike core is always the conductor.

Autonomous

Core-driven routing

The spike core evaluates every query against its living memory and automatically decides whether to answer on its own, call a local model, or reach for a frontier API — picking the most capable path without you having to think about it.

  • Zero configuration — works out of the box
  • Core weighs confidence, context sensitivity, and task complexity
  • Falls back to local-only if connectivity is absent
User-configured

Your routing rules

You define exactly which queries go where. Pin sensitive topics to local-only, always route code tasks to a specific model, or set hard caps on what can leave your infrastructure — the core enforces your rules on every call.

  • Per-topic, per-workspace, or per-user routing policies
  • Explicit allow-lists for cloud model access
  • Your rules are learned as policy memory — they persist
⚡

Voltaic Spike Core

Reasons from Living Memory — then routes via autonomous judgment or your configured rules.

On your metal
Local models

Open models running on-premise — sensitive context never leaves your hardware.

Frontier
Cloud models

Frontier APIs for raw scale — the core controls what's shared and when.

Orchestrated answer

Output flows back through the spike core and its memory — grounded in your context, always. And the core learns from every call.

🤖

Autonomous by default

Out of the box, Voltaic routes with no setup. The spiking core judges each query against its memory and picks the best path — local, cloud, or self — automatically.

🎛️

Configurable when it matters

Need hard guarantees? Lock sensitive topics to local-only, restrict cloud access by workspace, or set per-user policies. The core enforces your rules, every time.

🧬

Fused, not forwarded

External answers come back conditioned on what Voltaic knows about you — and every orchestrated call sharpens its memory. A truly orchestrated solution, not a proxy.

Proof, on the bench

Tested against the best. The gap is decisive.

On public continual-learning benchmarks, with no replay buffer, VOLTAIC retains 2.6×–4.2× more knowledge than dense 7B baselines as new domains arrive — and collapses forgetting to 0.04% on vision tasks when paired with HULL. Every claim is backed by 647 automated tests, 100% passing.

647/647tests · 100% passing
0.04%forgetting · CORe50-NI + HULL
4.2×vs Mistral-7B · ImageNet-1k
77.3%accuracy · Split ImageNet-1k
VOLTAIC Dense 7B baselines (Mistral-7B · Qwen2.5-7B) rehearsal-free · no replay buffer · CLBench v10
CORe50-NI · forgetting collapse with HULL
Standard model
1.20%
VOLTAIC + HULL
0.04%

Lower is better. HULL also lifted average accuracy 95.01% → 95.23% simultaneously.

CLBench v10 · no-replay sequential protocol · Single A100 (GPU 3) · 2026-06-30, seed 0. VOLTAIC retains 15.74% vs 6.14% / 5.91% on Split CIFAR-100 (2.6×) and 20.87% vs 4.93% / 4.90% on Split ImageNet-1k (4.2×). Vision tests: 8× A100 · DINOv2 features · run 2026-06-14.

The journey

Where we are & where we're going.

From a proof-of-concept spiking model to embodied intelligence on neuromorphic hardware.

  1. ✓
    Shipped

    The spiking core

    A from-scratch SNN language model — binary spikes, LIF neurons, STDP attention — trained on real language data and proven to learn. 647 tests green.

  2. ✓
    Shipped

    Memory that holds — and sharpens

    HSC v2 "Living Memory" + MemoryForge + HULL (Accumulative Hull Neuron Model). Forgetting collapses to 0.04% on CORe50-NI, accuracy improves simultaneously. Head-to-head against Mistral-7B and Qwen2.5-7B with no replay buffer: VOLTAIC retains 2.6–4.2× more knowledge on the hardest class-incremental tasks.

  3. ●
    We are here

    Validated, live & scaling

    The spiking core is proven on public benchmarks (CLBench v10), running live as a multimodal app — chat, teach, voice, camera-gesture learning, external model orchestration. 647 automated tests, 100% passing. The architecture is locked; we are now scaling the model and hardening the platform for production deployments.

Where we're going

Robots that learn on the job.

A frozen model can't adapt to a new factory line, a new tool, or a new home. A Voltaic brain can. By pairing the spiking core with the Accumulative Hull Neuron Model, the robot keeps a geometric "stability hull" around skills it has mastered — so it can pick up something new without unlearning how to walk.

  • Teach by demonstration — hands, voice, and vision in one brain.
  • Learns in the field, not in a retraining run.
  • Protected core skills stay protected, provably.
  • Built to run on low-power neuromorphic hardware.
Build the body with us →

Most models are frozen.
Ours is alive.

A model that learns new skills live, from any sense, without forgetting — and can explain itself — running on a single GPU. That's Voltaic. Let's give it a body.

hello@voltaic.live Back to top ↑