Compact Node
< 1 kW
- Apertus 8B
- gpt-oss-20b
- Qwen 27B INT4
Pilot, local RAG, one principal or small team.
Hardware · model fit
Memory first. Then throughput, power and support.
The smallest machine that holds a model is not automatically the machine that serves your team. We separate weight residency from context, concurrency, latency and operational headroom.
Official specificationsPlanning calculation
Does the model fit?
Choose a model and the aggregate accelerator-memory class. The result is a planning gate, never a performance promise.
Planning result
Select a model and memory class.
Reference components, not anonymous tiers
These specifications anchor the architecture. Final systems add CPU, RAM, storage, fabric, redundancy, warranty and serving validation.
96 GBRTX PRO 6000 Blackwell · ECC GDDR7 · 600 W
128 GBDGX Spark · coherent unified memory · proof-node class
141 GBH200 SXM · HBM3e · enterprise fabric
288 GBMI355X · HBM3e · 1.4 kW accelerator
A DGX Spark can hold large quantised weights in unified memory, but that does not make it equivalent to a high-bandwidth datacentre accelerator for latency or concurrency.
Four delivery envelopes
Pro Node
0.6–1.5 kW
- Kimi Linear INT4
- Apertus 70B INT4
- gpt-oss-120b validated path
The normal starting point for serious local production.
Vault
2–8 kW
- larger multimodal models
- higher concurrency
- redundant storage and service paths
Multi-user archive system.
Cell
8–40 kW
- Kimi K2.5 validated INT4
- DeepSeek V3 candidates
- separate development and production
Regulated team and large-model system.
Sovereign Rack
50–200+ kW
- Kimi K2 and K3 official serving envelopes
- K3 starts at 8× GB300 or MI350X-class; production may be multi-node
- liquid cooling and heat reuse
A facility engagement, never an appliance. Exact power and performance remain site- and benchmark-specific.
What the calculator deliberately does not promise
Context headroom
Longer context increases KV-cache and runtime demand. A weight fit at 2K can fail at 128K.
Concurrency
One fast user and twenty simultaneous users are different architectures even when the model is identical.
Quality
Hardware fit says nothing about groundedness, language quality or refusal behaviour on your archive.
Procurement rule: no named component becomes a client configuration until current Swiss availability, warranty path, lead time, exact chassis fit and benchmark results are recorded.