All local H3 guides
Performance / Local H3

MiniMax H3 local performance benchmarks

Real MiniMax H3 generation times on B300, H200, H100, dual RTX 5090, DGX Spark, L40S, and a 16 GB RTX 4090 Laptop, with test differences explained.

Reviewed
2026-08-09
Reading time
7 min
Evidence
1 official / 3 community
Short answer

Measured H3 speed ranges from seconds on large multi-GPU servers to minutes on offloaded workstations. SGLang reports about 13 seconds for a five-second 1344 x 768 request on four H100s and about 74 seconds on four H200s under a different measured path. Its dual RTX 5090 50-step run took about 560 seconds. Community low-memory results range from roughly three minutes for a five-second 960 x 540 ComfyUI clip to about nine minutes for a 5.17-second 768 x 576 DGX Spark run.

01

Read the workload before the number

A generation-time claim is incomplete without its workload. At minimum, record the output dimensions, exact frame count, denoising steps, task family, reference inputs, weight precision, attention backend, offload policy, warm or cold state, and whether model loading and decode are included.

02

Published end-to-end results

Hardware and runtimeWorkloadReported time
8 x B300 / SGLang BF16 FL2VA5.17 s, 1344 x 768, 50 steps, warm19.04 s latency
4 x H200 / SGLang Ulysses45 s, 1344 x 768, 50 steps, warm resolution74.38 s E2E
4 x H100 / SGLang TP2 + Ulysses2Published topology comparison13.25 s pipeline latency
2 x RTX 5090 / SGLang offload5 s, 1344 x 768, 50 steps559.67 s E2E
DGX Spark / Diffusers INT8 offload5.17 s, 768 x 576, 10 grid pointsabout 543 s optimized average
L40S / optimized raw weights10 s, 960 x 544about 7.2 min
RTX 4090 Laptop / ComfyUI repack5 s, 960 x 540, 20 steps182 s E2E
03

The largest performance levers

The L40S report is a useful example: keeping the text encoder resident starved the denoiser of memory, while streaming the video VAE leaf by leaf made decode much slower than necessary. Better component placement mattered more than a nominally faster card.

  • Canvas size and frame count: video token count dominates quickly.
  • Component residency: moving the text encoder away before denoising can free enough VRAM to stop transformer blocks streaming every step.
  • Video VAE placement: decode can dominate a run when its modules repeatedly cross the bus.
  • Parallel topology: tensor, Ulysses, Ring, and FSDP trade memory against communication and latency.
  • Approximation: INT8, FP8, few-step LoRAs, caching, and alternate attention need output-quality validation.
04

Build a benchmark that answers your decision

Use three fixed prompts: one text-only motion shot, one first-frame identity shot, and one reference-heavy shot. Run a preview canvas and the 768p target. Repeat a warm request at least three times, report the median, and inspect the same frames and audio timestamps.

  • Separate load, encode, denoise, video decode, audio decode, and mux time.
  • Record peak VRAM and host RAM.
  • Publish the model revision and all non-default flags.
  • Keep quality-changing optimizations in separate rows.
Frequently asked questions
How fast is MiniMax H3 locally?

It varies by more than an order of magnitude. Multi-GPU datacenter recipes can finish a five-second request in seconds or around a minute, while offloaded workstation paths commonly take several minutes.

Why are MiniMax H3 benchmark results so different?

Hardware, precision, resolution, frames, denoising steps, checkpoint family, references, attention backend, offloading, warmup, and whether timing includes decode all change the result.

How long does MiniMax H3 take on two RTX 5090 cards?

SGLang reports 559.67 seconds end to end for its verified 50-step, five-second, 1344 x 768 request using TP2 and layerwise offload.

Can I compare a five-step Turbo result with a 50-step BF16 result?

Only as separate operating points. A few-step LoRA, quantized checkpoint, or approximate attention path may change quality, so it is not a direct hardware speed comparison.

Sources and evidence

Read the original material

Community figures describe the named machine and workflow only. They are not MiniMax performance guarantees.

  1. Official documentationSGLang H3 benchmark tables
  2. Community benchmarkDGX Spark benchmark and optimization
  3. Community benchmarkRTX 4090 Laptop first run
  4. Community benchmarkL40S raw-weight benchmark