Open models at full precision, and the memory they need.
Running an open model at full 16-bit precision takes about 2 GB of GPU memory per billion parameters for the weights alone: Qwen3.5 397B needs about 794 GB in BF16, which one Wick Ultra's 1,152 GB holds.
Several frontier models are released below 16-bit, so their full precision is the format their makers published: Kimi K2.5 ships native INT4 weights and DeepSeek V4 Pro mixes FP4 experts with mostly FP8 other weights.
From $656,000Includes delivery, the outdoor unit and the first year of service.
What each model was released in
The released checkpoint is the reference for quality. Converting it up to 16-bit adds memory without adding accuracy; converting it down is a separate quantization step.
| Model | Parameters | Released as | 16-bit weights | Smallest Wick at 16-bit |
|---|---|---|---|---|
| Qwen3.5 397B | 397 billion | BF16 configuration; official FP8 variant available | 794 GB | Wick Ultra |
| GLM-5 | 744 billion | BF16 configuration; official FP8 variant available | 1,488 GB | Wick Cluster node |
| DeepSeek V4 Pro | 1.6 trillion | Mixed FP4 expert weights and mostly FP8 other weights | 3,200 GB | More than one Cluster node |
| Kimi K2.5 | 1 trillion | Native INT4 weight quantization, as in Kimi K2 Thinking | 2,000 GB | Wick Cluster node |
| Kimi K3 | 2.8 trillion | Native MXFP4 weights with MXFP8 activations | 5,600 GB | More than one Cluster node |
BF16, FP8 and 4-bit
BF16 and FP16 store each weight in 2 bytes, FP8 and INT8 in 1 byte, and FP4, INT4 and MXFP4 in half a byte, plus small scale factors. Each step down halves the weight memory.
NVIDIA's Rubin GPU has hardware for NVFP4, a 4-bit format: NVIDIA rates it at 50 petaflops for inference per GPU (sparse) and 35 for training (dense). Those are NVIDIA's specifications, not Wick measurements.
GPU memory in each Wick
16-bit capacity is GPU memory divided by two bytes per parameter, before context.
| System | GPU memory | 16-bit up to | 8-bit up to | 4-bit up to |
|---|---|---|---|---|
| Wick | 288 GB | About 144 billion | About 288 billion | About 576 billion |
| Wick Pro | 576 GB | About 288 billion | About 576 billion | About 1.15 trillion |
| Wick Ultra | 1,152 GB | About 576 billion | About 1.152 trillion | About 2.3 trillion |
| Wick Cluster node | 2,304 GB | About 1.152 trillion | About 2.304 trillion | About 4.6 trillion |
Questions
How much VRAM does a model need in BF16?
About 2 GB per billion parameters for the weights: 1,488 GB for GLM-5's 744 billion, and more for context. GLM-5 in BF16 fits one Wick Cluster node's 2,304 GB, and its official FP8 release, 744 GB, fits Wick Ultra.
Can DeepSeek V4 Pro run at full precision on one machine?
Its released mixed FP4 and FP8 checkpoint is its full precision. The uniform 4-bit floor is 800 GB, within Wick Ultra's 1,152 GB; the released checkpoint needs somewhat more, which Wick confirms for your Wick in your build review.
BF16 or FP8: which should a team serve?
Where a maker publishes an official FP8 release, as Qwen and GLM do, serving it halves the memory against BF16. Keep BF16 where you fine-tune or need the reference outputs.
Sources
- Qwen3.5 397B model card on Hugging Face, observed October 27, 2026
- GLM-5 model card on Hugging Face, observed October 27, 2026
- DeepSeek V4 Pro model card on Hugging Face, observed October 27, 2026
- Kimi K2.5 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.