SparkSE

An independent AI provider, in Sweden.

SparkSE is an independent provider serving open-weight models on hardware it owns. The weights are resident, the machines are in Sweden, and requests are answered on the Tenzro network.

01Models

Open weights, resident on the machines

Muse Glimmer

Reasoning, served with a speculative drafter alongside it.

Qwen 3.8

General and code work. The default for most requests.

Gemma 4

A third family, so no single lineage is the only option.

More to follow. Each family earns its place by being good at something the others are not.

02Hardware

DGX-class inference

NVIDIA DGX Spark — Grace Blackwell silicon with a unified memory pool the CPU and GPU address as one. A 30-billion-parameter model loads whole on a single unit and stays resident, and units are added as capacity is needed.

PlatformNVIDIA DGX Spark — GB10 Grace Blackwell
Memory128 GB unified LPDDR5x at 273 GB/s, per unit
LocationSweden
IdentityEach machine's own TPM — assurance: tpm, endorsed
NetworkAnswers on the Tenzro network. Furcate overlay, tier platform
Reached throughA passthrough gateway in Iceland holding no key. It forwards by server name without decrypting; the certificate stays in Sweden
03Energy

Each unit runs off one power supply

No hall, no substation, no cooling plant. Capacity grows a unit at a time, and a unit is a desk and a socket.

Peak draw, per unit240 W — the whole system, from one external supply
Where it goes140 W to the GB10 processor, 100 W for everything else
Idle40.9 W at the module
ScalingLinear. Another unit is another 240 W — no step change in cooling or supply
CoolingAir. No water, no chillers, no evaporative loss
Footprint150 × 150 × 50.5 mm, 1.2 kg

Capacity at this scale sits where the people using it already are — an office, a lab, a home — on power and cooling a building already has, and arrives in increments a community can own.

04Access

Independent, and reproducible

SparkSE is owned and run by its operator. It uses Furcate to run the machines and the Tenzro network to answer requests — both are open, so anyone can stand up a provider the same way, and nothing here is a permission somebody granted.