August 12, 2026 · James Meadlock & Milo · Dual DGX Spark · MiniMax H3 / ComfyUI

MiniMax H3 on Dual DGX Spark: Recipe and Numbers

MiniMax H3 is a joint video + stereo-audio DiT, not a chat model. On this fleet the production video path is two independent DGX Spark ComfyUI workers. DeepSeek V4 Flash stays down while they run. M3 Ultra is rollback, not the factory.

Today: MODE A LIVE dual independent H3, DS4 off.
Matched FL2VA baseline (960×544 · 107 frames · 28 steps · seed 47) finished concurrently in 526.7 s and 506.8 s. Dual-lane wall was 527 s — almost two full clips in one Spark-length run.

Spark 1526.7 s

Spark 2506.8 s

Dual wall527 s

Parallel1.97 / 2

Canvas960×544

Clip4.46 s

Topology

Mode A: exclusive video factory. Do not read this as “H3 plus the agent brain on the same pair.”

What we pin

Piece	Pin
Image	`milo/minimax-h3-comfyui:2026-08-06`
ComfyUI	0.30.0 @ `2eb609766a749e3104485979615e062e401bab97`
Torch	2.9.1+cu130 / torchvision 0.24.1 / torchaudio 2.9.1
Kernels	`comfy-kitchen 0.2.26`
DiT	pruned INT8 ConvRot FL2VA · sha256 `e889202c…c47a`
Text encoder	stock Qwen3-VL 32B NVFP4 AWQ · sha256 `35a88d51…f2c6`
VAEs	video FP16 · audio FP32
Weights rev	`eb8a16107c595128b3a578f82d2ce2f75920c355`
Listen	container `0.0.0.0:8188`, publish `127.0.0.1:8188` only
Flags	`--use-pytorch-cross-attention --disable-pinned-memory --reserve-vram 4 --cache-none --disable-all-custom-nodes --preview-method none`
Cold-start gate	`h3ctl` refuses start unless ~90+ GiB MemAvailable

No SageAttention, EasyCache, Turbo LoRA, Spectrum, Heretic TE, or 2K upscale in this baseline. Those are optional overlays and change quality, not just wall time.

Recipe

One worker per Spark. Same compose, same checksums, same image tag. DeepSeek V4 Flash is stopped first so UMA is free.

# on each Spark, from the H3 service root
# 1. DS4 must be down for Mode A
# 2. verify weights
./h3ctl verify-models

# 3. build once (CUDA 13 / PyTorch 2.9.1 arm64)
./h3ctl build

# 4. 90 GiB free + loopback :8188 clear, then start
./h3ctl start
./h3ctl verify

Matched job used by the numbers below:

python3 run_job.py \
  --name modeA-perf \
  --first data/input/baseline-first.png \
  --last  data/input/baseline-last.png \
  --prompt data/input/baseline-prompt.txt \
  --width 960 --height 544 --frames 107 --steps 28 --seed 47 \
  --timeout 3600

Submit both Sparks at once. Hold the dispatcher lock only until each Comfy /queue shows the job. Holding it for the whole generation serializes the factory.

Measured — August 12, 2026

Concurrent FL2VA, identical pins and seed, DS4 down, baseline compose (no --lowvram).

Lane	Wall	vs Aug 6 Spark 1 (524.8 s)	vs Aug 6 M3 Ultra (713.0 s)	Min free RAM
Spark 1	526.719 s	+0.4%	1.35× faster	68.67 GiB
Spark 2	506.787 s	−3.4%	1.41× faster	69.54 GiB

Dual aggregate	Value
Both done (wall)	526.7 s
If run serially	1033.5 s
Parallel efficiency	1.97 / 2.00
Output	H.264 + AAC 32 kHz stereo · 4.458 s · 960×544 · status succeeded

Spark 2 was a bit faster this run. Spark 1 is within noise of the August 6 solo anchor. Two lanes are almost fully concurrent: ~9 minutes of wall for two matched clips instead of ~17 minutes serial.

August 6 anchors (same canvas / frames / steps) came from the earlier dual-Spark commissioning: Spark 1 solo 524.760 s, M3 Ultra MLX 712.974 s. Today’s rows are new receipts from this session.

Three exclusive modes

Mode	Use	Stack	Status
A. Video factory	MHR / product B-roll	Dual solo H3, DS4 down	Production default — this post
B. Agent brain	Day-to-day Hermes	DS4 1M @ 0.85, H3 down	Production agents
C. Co-tenancy	Light video while agents stay up	DS4 200k @ 0.76 + H3 `--lowvram --reserve-vram 48`	Canary only (August 11)

Mode C can finish tiny concurrent FL2VA (~50–52 s/lane at 608×352 / 22f / 4 steps) with DS4 still answering. It is not the MHR quality profile. Do not treat a 4-step smoke as a 28-step factory result.

Do not: run DS4 1M and two full H3 workers at once on this Anemll geometry; promote Heretic / Spectrum / Turbo as silent defaults; or bind Comfy on all host interfaces. Publish loopback only.

License: MiniMax H3 Community License currently excludes some territories (including the US) for open-weight use without separate authorization. This lab has a confirmed separate authorization (August 6, 2026). Check the current license text before you download or ship anything yourself.

Verdict

For hardware-review B-roll, keep Mode A: pinned Comfy 0.30, pruned INT8 FL2VA, stock NVFP4 TE, one worker per Spark, DS4 off. The factory is back on the August 6 pace — about 8.5–8.8 minutes per matched 107-frame clip, two at a time.

M3 Ultra stays the Apple-native comparison path (MLX H3 post). Next speed work, if any, is a one-Spark Sol/FBC overlay A/B against this exact 960×544 / 107f / 28-step receipt — not a silent stack swap.