Back to Verkko Robotics

Verkko Robotics · Unveiled 18 June 2026

VOLTAIC

VARIABLE OUTPUT LANGUAGE TRANSFORMER WITH ADAPTIVE INTEGRATED COMPUTATION

7B

Spiking parameters

88%

Network sparsity

~1.2%

Forgetting rate · Core50

1/50th

Cost vs leading LLMs

A 7B-parameter spiking language model that learns continuously, never forgets, and runs on your hardware.

VOLTAIC is engineered for three things at once: continuous learning without catastrophic forgetting, extreme energy efficiency via 88% sparsity, and sovereign edge deployment — no cloud required.

Visit voltaic.live
Four translucent layers stacked in 3D. Each layer is a node mesh, connected vertically by light traces.

Why it exists

Conventional AI was not built for the field.

Today’s dominant models depend on two assumptions: that compute is cheap and centralised, and that learning happens once, in a training run, far from where the model is used.

Both assumptions break the moment AI leaves the browser. Edge environments — vehicles, factories, infrastructure, devices — cannot afford the energy bill, the latency, or the data exposure of a cloud round-trip. They also cannot afford a model that is frozen on the day it shipped.

VOLTAIC is built for the world after that assumption breaks.

Energy by orders of magnitude

Sparse spike activity means most weights contribute zero energy at inference time. The economics of edge AI change when the network is silent more often than not.

Continuous learning, not just inference

Most foundation models are frozen after training. VOLTAIC adapts in place — incorporating new knowledge online while preserving prior capability.

Sovereign by architecture

No cloud dependency. Data does not leave the device unless an operator chooses for it to. Privacy and resilience follow from the system design, not from policy.

Architecture

One platform.
Four moving parts.

VOLTAIC is a fully integrated spiking language model — four co-designed components that each solve a hard constraint of edge intelligence: computation, memory, retention, and deployment.

Spiking Inference Core

Computation

Every layer of the network operates on binary spikes over discrete timesteps — the same primitive used by biological neurons. Most of the network is silent at any moment; only the part that is needed fires.

Living Memory

Continuous learning

A memory layer in which persistence is earned, not assumed. New domains are absorbed without erasing what came before. The model stays sharp across sequential exposure to different bodies of knowledge.

Verkko Equilibrium Engine (VEE)

Retention control

A principled allocator that decides how attention is balanced across past and present knowledge. Replay schedules emerge from a stable equilibrium — no manual tuning, no catastrophic forgetting.

Sovereign Edge Runtime

Deployment

Runs on local hardware. No cloud round-trip required for inference or for incremental learning. Designed for environments where latency, power, or data sovereignty are non-negotiable.

Core IP

Patent pending

The Accumulative Hull Neuron Model.

Continual learning fails in a specific way. A model that absorbs a new domain tends to lose the last one, and the usual defences — replay buffers, regularisation on individual weights, frozen layers — trade away either capacity or the ability to keep learning at all.

AHNM is the subsystem inside VOLTAIC that resolves this. It governs stability at the level of neuron activation geometry rather than at the level of individual synapses, which lets the network hold on to durable knowledge while remaining plastic at the periphery. It is what makes Living Memory and the Verkko Equilibrium Engine work in practice.

Knowledge that matters is protected structurally. Everything else stays free to change.

~1.2%

Forgetting rate

Measured on Core50 across sequential exposure, where conventional fine-tuning degrades sharply.

0.71×

Average drift

Across fifteen sequential domains. Thirteen of the fifteen showed backward transfer — earlier domains improved.

77%

Split ImageNet-1K

Fourteen points above the prior published record for this continual-learning benchmark.

Trajectory

Where VOLTAIC is going.

Validated across 15 sequential domains. Average drift 0.71× — meaning on average, domains held or improved. 13 out of 15 showed backward transfer. Now scaling to full deployment.

Now

7B params · 88% sparsity

Spiking language model with Living Memory validated on 15 sequential domains. Forgetting rate ~1.2% on Core50. 77% on Split ImageNet-1K — 14 pts above prior record.

Q4 2026

Release

Production deployment in industrial, sovereign infrastructure, and field environments where cloud AI cannot operate.

The destination

70B · 95% sparsity

Always-on personal intelligence that lives on your device, knows your life, and belongs to you — architecturally, not legally.

IP & partnerships

Held by Verkko Robotics.

VOLTAIC is the core IP of Verkko Robotics Ltd, developed at our Pisa R&D division and operated under the Verkko Research Programme.

The product is live at voltaic.live. Commercial deployment and licensing runs through Verkko.ai. Joint research and reference pilots are available to selected partners.

Engagement paths

Reference pilots

Industrial, environmental, or sovereign deployments validated in the field.

Joint research

Institutions and labs collaborating on neuromorphic, plasticity, or embodied directions.

Licensing & strategic partners

For organisations integrating VOLTAIC into proprietary stacks under an IP-safe arrangement.

Open a conversation