Catalog
HS-LU / Tiers

HS-LU-01 / Catalog

Three workstations,
plus one rack server.

Every Helmstone SKU is positioned for a class of job — single-GPU fine-tuning, multi-day inference, frontier-class training, or rack-scale continuous inference. Pick the class; the catalog returns the starting price, the headline spec sheet, and the current commercial path. Proposed systems are marked pending rather than presented as fully published validation.

HS-LU-02 / Tier catalog

The four systems, side-by-side.

Each card is a catalog summary: SKU, tier class, headline accelerator, dataplane memory + storage, starting price, and intended workload. Pick the closest fit and route to engineering.

HS-LU-03 / Picking a tier

How a buyer picks one.

The slate is intentionally short — four SKUs that cover the workload classes a research team or a lab actually purchases. SKU breadth is not the moat; picking the right one for the job is.

Local development & 8B–70B fine-tune

Helmstone W1. The single RTX 5090 covers consumer-grade fine-tune and scientific sim; 48h FP16 inference burn-in.

Multi-day 70B–120B inference

Helmstone W2 (dual Pro 6000). 192 GB combined VRAM is the threshold where KV-cache pressure stops throttling the run.

Frontier training to 72 hours

Helmstone W4. Four B200s with liquid cooling so the worst hour of the 48h burn-in is not the throttle event.

Overnight sweeps & continuous fleets

Helmstone RS. Rack-mounted for teams running concurrent jobs in the same validated thermal envelope.

See where on-site coverage lands →

Five Tier-1 US metros carry on-site engineers on the calendar; outside those metros the worldwide Tier 2 contract still covers. Open the coverage map at /service-area.

HS-LU / Engage

Not sure which tier fits?
Start with the workload.

Write to us with the model class, the expected job duration per run, and the service metro you are in. We will return a configuration, a benchmark target, and a quote — signed by an engineer.

What to include

  • •Model class and size (parameters, tokenizer, context length)
  • •Expected job duration per run, and runs per week
  • •Site location and preferred service metro
  • •Procurement vehicle (PO, grant, capital)