Arao AI
Back to search results

MiniMax-H3 · Online dynamic FP8

Start a discussion
NVIDIA DGX Spark128GBvLLM-Omni vLLM-Omni 0.1.dev2381+g310b4b477L0 Self-reported

This page aggregates 1 real-world runs of MiniMax-H3 (Online dynamic FP8) on NVIDIA DGX Spark with vLLM-Omni, contributed by 1 independent source platforms; metrics are averages of published measurements.

80.58 s

Gen time

Per image / per clip

89.17 GB

VRAM

VRAM usage

L0 Self-reported

1

Measured runs

1

Independent sources

GitHub

Source platforms

6 days ago

Last verified

Performance

No published records for this metric under the same model + hardware yet

Core figures

Gen time (avg)
80.58 s
VRAM (avg)
89.17 GB
Power draw
— W
Output
768x448 · 56 frames · 20 steps · 24fps

Configuration

Member-level fields are taken from the most recent run

Model
MiniMax-H3
Quantization
Online dynamic FP8
Framework
vLLM-Omni
Version
vLLM-Omni 0.1.dev2381+g310b4b477
Batch size
1
Output spec
768x448 · 56 frames · 20 steps · 24fps

Hardware

Nominal and measured figures are shown side by side; whether it runs is the reader's call

GPU
NVIDIA DGX Spark
Nominal VRAM
128 GB
Measured VRAM (avg)
89.17 GB
OS
Linux ARM64
Driver
580.173.02
CUDA
13.0

Get started

Weights for this deployment

Original model resources

Base model resources may not match this quantization. Use the verified deployment weights above when available.

Startup command

Command from the latest benchmark run
H3_DIFFUSION_ATTENTION_BACKEND=CUDNN_ATTN H3_EXECUTION_MODE=compile H3_CACHE_BACKEND=cache_dit H3_CACHE_CONFIG='{"Fn_compute_blocks":1,"Bn_compute_blocks":0,"max_warmup_steps":4,"max_cached_steps":-1,"residual_diff_threshold":0.10,"max_continuous_cached_steps":1,"enable_taylorseer":false}' docker compose up -d

Deployment guide

https://github.com/joeynyc/MiniMax-H3-DGX-Spark#quick-start

Sources & evidence

1 measured records in total, each traceable to its original source

  1. L0 Self-reportedGitHubOriginal link Verified on 2026-09-27

    vLLM-Omni 0.1.dev2381+g310b4b477 · Linux ARM64 · CUDA 13.0

    80.58 s

    Gen time

    89.17 GB

    VRAM

    — W

    Power draw

    单DGX Spark，balanced（cuDNN+compile+Cache-DiT0.10）：请求2秒、20步，输出56帧/768×448/24fps。warm两次client mean80.579s；89.1659GiB是模型加载统一内存，非峰值，整机peak113.09GiB。排除首次编译；缓存非无损，full-compute111.373s。