Skip to content
localmodel.run

video model · mochi · Linux

Can I run Mochi 1 on Nvidia DGX Spark (128GB)?

Compatibility verdict VRAM check
Yes, it runs fast on this GPU

Yes. Mochi 1 runs on Nvidia DGX Spark (128GB) at fp8 + offload (~20 GB of ~112 GB usable).

Needs ~20 GB Device usable ~112 GB

Runs at fp8 + offload using ~20 GB of ~112 GB usable.

Peak VRAM
~20 GB
Usable on device
~112 GB
Device memory
128 GB
Quant
fp8 + offload
Share on X Share on Reddit

How to run it

Use ComfyUI or Diffusers at fp8 + offload. The big text encoder is loaded to encode your prompt, then offloaded before generation, which is why peak VRAM stays near the backbone size rather than the sum of every file.

Model mochi
Type
video (DIT)
Parameters
10B
Peak VRAM
~20 GB at fp8 + offload
Resolution
480×848
License
Apache-2.0
Full Mochi 1 requirements →
Device Linux
Memory
128 GB unified
Usable for weights
~112 GB
Power draw
~140 W
Best runtime
llama.cpp (CUDA) / Ollama (CUDA)
Best models for Nvidia DGX Spark (128GB) →

You could also run

Run Mochi 1 on other hardware

Nvidia GeForce RTX 4090 (24GB)Nvidia GeForce RTX 3090 (24GB)Nvidia GeForce RTX 5090 (32GB)

FAQ

Can Nvidia DGX Spark (128GB) run Mochi 1?

Yes. Mochi 1 runs on Nvidia DGX Spark (128GB) at fp8 + offload (~20 GB of ~112 GB usable).

How much VRAM does Mochi 1 need?

Nvidia DGX Spark (128GB) has room to spare. At fp8 + offload the realistic peak is ~20 GB of VRAM, versus ~60 GB with every component kept resident (no offload). With aggressive CPU offload it drops to ~18 GB, much slower.

What do I use to run Mochi 1 locally?

Mochi 1 runs in ComfyUI or Diffusers. It loads as a video diffusion checkpoint plus its text encoder and VAE, not a single chat command.

Sources

VRAM figures are sourced peak-usage anchors at the noted quant; catalog updated 2026-11-01. See methodology.