LOG 00:00:01
MODEL: KIMI K3 / 2.8T MOE
TOPOLOGY: 1 HUB / N EDGE
STATUS: FIT-CHECKED
N · FULL PRECISION · MXFP4
GB200 NVL72 RACK · 72 GPU
All 896 Experts Resident
896 EXPERTS · 16 ACTIVE/TOKEN
CAPACITY: ~13.8 TB HBM3e
MODEL FOOTPRINT: 1.56 TB (Q8)
HEADROOM: ~12.2 TB FOR KV / BATCH
ROLE: VERIFIER · AUTHORITY
Hub site footprint — 5 containers total (1 compute rack + 4 below)
Cooling / Power Dist.
Liquid CDU · switchgear
Battery ~2 MWh
LFP · overnight + cloud day
Solar Transit 1 / 2
~720 panels packed
Solar Transit 2 / 2
~720 panels packed
~1,440 panels total · ~120kW array · empties out once field is deployed on-site
When connected
The edge node drafts the response fast and cheap, and the hub checks it over the link in batches — so you get output close to hub quality without sending every single word back and forth.
Sharing expertise
Each edge keeps the handful of experts it uses constantly close by, and asks the hub for the rare ones only when it actually needs them — no need to carry all 896 around locally.
When the link drops
The edge just keeps going on its own. Answers get a bit rougher without the hub to lean on, but the site never goes dark.
Edge sites — 2 containers each (compute rack + solar transit below)
EDGE NODE — SITE A
N
1-bit Quant
594GB · 8×H200 · standalone-capable
Solar Transit
~100 panels · ~38kW array
EDGE NODE — SITE B
N
2-bit Quant
861GB · 8×H200 · standalone-capable
Solar Transit
~100 panels · ~38kW array
EDGE NODE — SITE C
N
1-bit Quant
594GB · 8×H100 · tight, low headroom
Solar Transit
~100 panels · ~38kW array
Deployed Solar Footprint — Once Panels Leave the Container