MiniMax H3 tool

MiniMax H3 system RAM requirements

Our two measured workloads reached nearly the same system-memory peak even though their workloads were roughly 40x apart.

Your machine

For example Intel Core i9-10850K, the bench CPU. It does not change any light.

Optional; the bench has 16. Under 8 only slows offloading.

32 means 32,768 MiB, the unit this page measures in.

Prefilled from the GPU; edit it for a variant board.

Optional. Recorded, never credited: no run here used swap.

The ComfyUI default. Turn it off if you start with --disable-pinned-memory; the system RAM light switches to the other threshold set.

RedmeasuredEngine v1.0.3

Not enough as configured. System RAM is the bottleneck.

GreenVRAMmeasured

This site completed 1344×768 for 124 frames at a peak of 11,649 MiB on a 12,288 MiB card. Green means a run finished in this band, not that headroom is left: the 639 MiB gap is what ComfyUI left after filling the card.

Actual
12,288 MiB
Threshold
12,288 MiB
RedSystem RAMmeasured

Peak system RAM reached 43,587 MiB on our test machine. A 32GB machine has 32,768 MiB in total, leaving a shortfall of about 10.6GB before the operating system takes its share.

Actual
32,768 MiB
Threshold
45,958 MiB
GreenCPU

Logical CPU count not given. CPU only affects offload speed here and never decides the overall light.

Fixes to try

  1. Start ComfyUI with --disable-pinned-memory. One RTX 3090 report dropped system RAM from 29.8GB to 7.5GB; an Ubuntu RTX 3060 report stopped whole-machine freezes with the same flag. Turn the switch off above to see your machine judged that way.--disable-pinned-memory
    saves: 22,835 MiBreported
  2. Swap the Qwen3-VL-32B text encoder for the Qwen3-VL-4B conversion. The report behind it (15.7GB down to 5.2GB) was written as VRAM, not system RAM, so treat the host-side saving as unmeasured.
    reported

Next

Hosted comparison

Nothing is charged and no sign-in is asked for. This records one anonymous intent event and expands the comparison — the page and URL stay exactly as they are.

Threshold table 1.0.3 · where each line comes from

Read the full guide

The short answer

Peak system RAM reached 43,587 MiB on our test machine. A 32GB machine has 32,768 MiB in total, leaving a shortfall of about 10.6GB before the operating system takes its share.

Two runs, 40x apart in workload

Both site-measured workloads used the shared environment listed under Test conditions, including the same official R2V template, Ref2VA pruned INT8 and a fixed seed.

MeasurementSmoke runFull run
Wall time55.6 s2,581.8 s
Peak VRAM11,591 MiB (94.3%)11,649 MiB (94.8%)
Peak system RAM43,176 MiB43,587 MiB

The workload increased by roughly 40x and wall time by roughly 46x, while peak system RAM increased by only 411 MiB, under 1%.

The smoke run rendered 512×288 for 22 frames; the full run rendered 1344×768 for 124 frames. That is 3,244,032 pixel-frames against 127,991,808, a factor of 39.5.

Why lowering resolution does not help

For these measured runs, the memory budget was determined by resident model weights and working buffers rather than output size. Lowering resolution bought time, not memory: the much smaller workload completed sooner without materially changing the system-RAM peak.

Do not treat two points as a curve. These are two measurements from the same configuration, not a basis for broader extrapolation. They do not establish behavior for other model variants, runs with audio enabled, or outputs large enough to exceed VRAM.

Test conditions

  • RTX 3060 12GB on physical GPU 0.
  • Proxmox VE VM 100 with 16 vCPU, 47.05 GiB RAM and no swap.
  • ComfyUI 0.31.0 and PyTorch 2.13.0+cu130.
  • Official R2V template with Ref2VA pruned INT8 and a fixed seed.
  • The smoke run rendered 512×288 for 22 frames at 24fps; the full run rendered 1344×768 for 124 frames.
  • Audio, Turbo LoRA and SageAttention were not enabled.
  • System memory was measured from cgroup memory.peak, supported by VmHWM and MemAvailable_min.

Where each line comes from

Every band the checker uses, with its lower bound, its light and the evidence behind it. Bands marked measured come from this site's own run records; bands marked reported come from first-hand posts with the hardware stated.

The same table, as JSON, sits inside this page and is the file the site's estimation engine reads; the checker is one implementation of it, the engine another. Pinned memory decides which system RAM ladder applies because one report moved a machine two lights by turning it off.

Frequently asked questions

Is 32GB of RAM enough to run MiniMax H3 locally?

Not on our test machine. Peak system RAM was 43,587 MiB, while a 32GB system has 32,768 MiB in total. That is a shortfall of about 10.6GB before the operating system takes its share.

Does lowering the resolution reduce MiniMax H3 memory use?

Barely. We ran the same workflow at two sizes about 40 times apart in workload. Wall time went from 55.6 seconds to 2,581.8 seconds, but peak system RAM moved only from 43,176 MiB to 43,587 MiB, a difference under one percent. Lowering resolution buys time, not memory.

How much system RAM should I have for MiniMax H3?

Our runs peaked just under 43GB on a machine with 47.05 GiB available, leaving very little headroom. Anything at or below 32GB was short on every run we measured.

Does MiniMax H3 need 64GB of RAM?

Not strictly. With pinned memory on, every completed run this site has recorded peaked between 42,060 and 43,910 MiB, so a 48GB machine finishes with a few gigabytes to spare and 64GB is the first common size with comfortable headroom. With pinned memory off, one RTX 3090 report dropped system RAM from 29.8GB to 7.5GB, which is why the checker above turns a 32GB machine green in that mode. That flag has not been measured on this site's own bench.

Selected results

A selection, not a firehose — reviewed results next to this site's own evidence.

Site presets

Community results

Community results are temporarily unavailable.