Hardware for a 1-trillion-parameter model, in your building.
A 1-trillion-parameter open model such as Kimi K2.5 needs about 500 GB of GPU memory for its weights at 4-bit and about 1,000 GB at 8-bit, before context.
Wick Pro's 576 GB holds the 4-bit weights, and Wick Ultra's 1,152 GB holds the 8-bit weights, or the 4-bit weights with about 652 GB left for context and users.
From $656,000Includes delivery, the outdoor unit and the first year of service.
Kimi K2.5 in each Wick
Kimi K2.5 is a mixture-of-experts model with a trillion parameters in total, every one of which sits in GPU memory while it runs.
| System | GPU memory | 4-bit (500 GB) | 8-bit (1,000 GB) | Price |
|---|---|---|---|---|
| Wick | 288 GB | Beyond one unit | Beyond one unit | $199,000 |
| Wick Pro | 576 GB | Holds it, 76 GB left | Beyond one unit | $339,000 |
| Wick Ultra | 1,152 GB | Holds it, 652 GB left | Holds it, 152 GB left | $656,000 |
| Wick Cluster node | 2,304 GB | Holds it, 1,804 GB left | Holds it, 1,304 GB left | $1,234,000 per node |
Why the room left matters
What stays free after the weights holds each conversation's context and each person served at once. On Wick Pro, 76 GB remain at 4-bit, enough for one researcher's work. On Wick Ultra, 652 GB remain, room for a team and long documents, or the 8-bit weights for higher fidelity.
Past a trillion
DeepSeek V4 Pro, at 1.6 trillion parameters, takes about 800 GB at 4-bit, within one Wick Ultra. Qwen3.8 2.4T and Kimi K3, at 1,200 GB and 1,400 GB, fit in one Wick Cluster node's 2,304 GB.
Two Wick Ultra units stack flush, but each keeps its own GPU memory and joins the other over fiber, so a stack adds a second computer rather than one larger memory.
The room it needs
Wick Pro runs on a 240 V, 50 A circuit, Wick Ultra on 240 V, 60 A circuit. Each sends its heat outside through its own outdoor unit. Wick reviews your room with you before you accept a price.
Questions
Can one GPU run a 1-trillion-parameter model?
Not from its own memory. One Rubin GPU's 288 GB holds 4-bit weights up to about 576 billion parameters, so a trillion needs at least two, as in Wick Pro.
Can I fine-tune a 1-trillion-parameter model on one Wick?
Not fully: full fine-tuning holds about 16 bytes per parameter, about 16,000 GB for a trillion. One Wick Ultra fully fine-tunes models up to about 72 billion parameters, and adapter methods reach larger ones.
What does it cost to own?
Wick Pro is $339,000 and Wick Ultra $656,000, list price in USD before tax. Every list price includes delivery, Wick OS and the first year of service. Wick, Wick Pro and Wick Ultra also include their own outdoor unit, fitted the same day. Wick Cluster's outdoor units or building-water connection are confirmed in your written price.
Sources
- Kimi K2.5 model card on Hugging Face, observed October 27, 2026
- DeepSeek V4 Pro model card on Hugging Face, observed October 27, 2026
- Kimi K3 model card on Hugging Face, observed October 27, 2026
- NVIDIA, Vera Rubin NVL72 specifications, per Rubin GPU, observed October 23, 2026
Wick figures are read from its published specifications and list prices. Checked October 27, 2026.