MiniMax H3 Local Setup Guide: ComfyUI, Models, VRAM and First Render

AI Overview

Can MiniMax H3 run locally?

Yes. MiniMax released H3 weights and ComfyUI-compatible packages, but a usable installation needs the correct diffusion model, text encoder, video VAE, audio VAE, workflow, and enough system memory for offloading.

How much VRAM does MiniMax H3 need?

There is no single requirement because precision, quantization, resolution, duration, and offloading change memory use. A 24 GB GPU is a practical target for pruned INT8 workflows; smaller cards need stronger quantization and patience.

Which MiniMax H3 model should I download?

Choose FL2VA for text, first-frame, last-frame, or image-led generation. Choose Ref2VA when the workflow consumes multiple reference images, videos, or audio. Never download the full repository blindly.

Is local H3 better than Seedance Agent?

Local H3 offers maximum graph and file control. Seedance Agent is better when you want H3 output without maintaining drivers, weights, VRAM workarounds, shot records, approvals, and partial reruns yourself.

Decide Whether Your Machine Is Ready

Check GPU memory, system RAM, and storage together

A MiniMax H3 local setup guide has to begin with capacity, not installation commands. The official MiniMax repository is close to 498 GB because it contains multiple pipelines and components. You usually need only one diffusion checkpoint, one matching text encoder, the video VAE, the audio VAE, and one workflow. Even a reduced stack can occupy tens of gigabytes before caches, temporary files, and output video are counted.

For a straightforward ComfyUI start, 24 GB of VRAM is a practical target for a pruned INT8 diffusion model plus an aggressively reduced text encoder with offloading. Twelve to sixteen gigabytes can work with community GGUF or other quantized routes, but setup becomes more sensitive and generation slower. Eight-gigabyte claims normally depend on heavy CPU/RAM offload and a specific runtime, so treat them as “can start” rather than “comfortable.” Keep generous system RAM and fast SSD space available; VRAM is not the only limit.

The hosted MiniMax H3 generator is the quickest control test. Run the same short idea there first. If the local result fails, you can separate prompt or model limitations from an installation problem.

A finished-output concept created for this guide. Face, clothing, crescent geometry, tool contact, sparks, and workshop layout form a demanding first local quality check.

Treat Apple Silicon as a separate path

The official integration index now points to an experimental Metal-native H3 route for Apple Silicon, but it is not the same operational path as the standard CUDA-and-ComfyUI setup. Do not copy NVIDIA wheel commands onto a Mac and expect them to work. Verify the chosen project’s supported chip, memory, workflow types, and current limitations, then use a small test. For dependable production on a Mac, a hosted route may be faster than debugging a young local port.

Choose the Correct Stack and Workflow

Start with ComfyUI unless you need library-level control

ComfyUI is the most approachable local route because current templates encode the model loader, text encoder, VAE, sampler, audio decode, and final video assembly. Update ComfyUI before importing an H3 template; older stable builds may not know the required node types. Begin with a core template and no optional custom nodes. Add acceleration, caching, interpolation, or upscaling only after one clean baseline render works.

A direct Diffusers or custom Python path is useful for service development and research, but it exposes more dependency and pipeline decisions. The simple snippet shown on a model hosting page may not represent the complete audio-video workflow. If your goal is one local clip, start from a maintained workflow rather than reconstructing every component from filenames.

Match FL2VA or Ref2VA to the job

FL2VA is the family used for text-to-video, image-to-video, and first/last-frame control. Ref2VA is a different checkpoint family for reference images, videos, and audio. Loading a Ref2VA workflow with FL2VA weights is not a harmless substitution: the graph expects a different conditioning route. The Ref2V versus FL2VA guide explains the decision in detail.

For a first test, choose FL2VA text-to-video or one-image I2V. They have fewer moving parts. Move to Ref2VA only when the baseline proves the runtime, VAE, and output chain are healthy.

Download and Place the Required Files

Build a manifest before downloading

Write down the exact workflow filename and every model filename it references. A practical manifest has four rows: diffusion model, text encoder, video VAE, and audio VAE. Record precision, file size, source repository, target directory, and checksum when one is published. This prevents two costly mistakes: downloading every available variant and pairing files that were not designed for the same graph.

For a current ComfyUI package, the directory pattern is:

ComfyUI/models/diffusion_models/<H3 diffusion checkpoint>
ComfyUI/models/text_encoders/<matching Qwen3-VL text encoder>
ComfyUI/models/vae/<MiniMax H3 video VAE>
ComfyUI/models/vae/<MiniMax H3 audio VAE>

The open-weight local pipeline produces 768p video; hosted 2K regeneration is a separate service. Do not download a file labeled 2K expecting the normal local graph to gain that output mode.

Verify filenames before launching

After each download, compare byte size and checksum, then confirm the filename in the workflow selector exactly matches the file on disk. Partial model downloads can still appear in a folder and fail later with confusing tensor or deserialization errors. Keep FL2VA and Ref2VA diffusion weights clearly named. If you test several quantizations, change one component at a time and save each working graph under a new versioned name.

A close finished frame reveals whether a local configuration preserves identity, hand contact, tool geometry, and the sculpture when camera distance changes.

Import the Workflow and Make a Baseline Render

Load first, then resolve red nodes

Open the official template in an updated ComfyUI installation. If nodes are red or unknown, stop and identify whether they belong to current core ComfyUI or a documented custom package. Do not install every node suggested by a manager. Restart after required packages change, then reopen the template and select the four manifest files in their correct loaders.

Set a short duration, conservative landscape resolution, standard sampler settings from the template, and one simple prompt. Avoid LoRAs, reference packs, upscalers, frame interpolation, and long multi-shot prompts. Your first acceptance criteria are mechanical: the prompt queues, model loads once, sampling advances, video and audio decode, the final MP4 plays, and a second run does not redownload or recompile unexpectedly.

Use one prompt that exposes common failures

Try a subject performing one physical action in a well-lit place with an audible event: “A blacksmith in an indigo shirt and tan apron strikes a crescent sculpture once, sparks arc left, then she lowers it toward a quench tank; bright workshop daylight, stable face and hands, clear hammer impact and steam.” This reveals identity, object geometry, contact, motion, light, and sound without becoming a full story.

MiniMax H3 first-and-last-frame output sample

A finished H3 output for checking whether the local graph produces real motion, a coherent transition, and a complete playable file rather than isolated frames.

Fix the Most Common Local Failures

Diagnose by the stage that failed

An out-of-memory error during text encoding points to the text encoder or its offload strategy; one during sampling points to the diffusion model, latent size, duration, attention implementation, or offload behavior; one during decode points to the video/audio VAE and available headroom. Record the last completed node instead of treating every crash as “not enough VRAM.” Reduce duration and resolution first, close competing GPU apps, and test a known-compatible quantization.

If a VAE will not load, verify that video and audio files are in models/vae, complete, and selected by the correct loader. If the final video has no sound, confirm the audio VAE decoded the audio portion of the joint latent and that the final video node muxed both streams. The H3 VAE and TensorRT speed guide helps separate decode bottlenecks from sampling speed.

Steam, water contact, tongs, face, and crescent shape create a harder motion-and-decode checkpoint after the simple hammer shot passes.

Add optimizations one at a time

Once the baseline works, duplicate the workflow before adding a Turbo LoRA, SageAttention, block caching, another quantization, or a custom VAE path. Change one variable and rerun the same seed and prompt. Record peak VRAM, wall time, output size, visible artifacts, and audio quality. Four-step acceleration can alter fast motion or sound; speed only counts when the clip still passes review.

Turn One Working Graph into a Production Workflow

Version the graph, environment, and output together

Save the working workflow JSON with a manifest containing ComfyUI commit, custom-node commits, Python, PyTorch, CUDA, GPU, model hashes, seed, dimensions, duration, and sampler settings. Export an API-format workflow separately if you plan to submit it through /prompt; the editor JSON and API JSON are not interchangeable. Keep previews, final MP4s, and logs under the same job identifier.

For longer stories, use short blocks and carry accepted endpoints forward. The multi-keyframe workflow covers intermediate visual anchors, while the resumable multishot workflow shows how checkpoints prevent a failure near the end from restarting the entire job.

MiniMax H3 reference-to-video output sample

A distinct finished H3 reference output for reviewing identity, reference influence, camera behavior, native audio, and the final muxed file after the basic FL2VA path is stable.

Choose Local H3 or Seedance Agent

Keep local control when it creates real value

Local H3 makes sense when model files must remain on your machine, you need custom nodes or experimental quantizations, you can maintain a GPU environment, and iteration time is acceptable. It gives you direct access to workflow JSON, seeds, intermediate states, and automation endpoints. That control is valuable for research and specialized pipelines, but creates an ongoing maintenance obligation.

Seedance Agent fits teams that want to use H3 without turning driver compatibility, storage, offloading, node versions, and render recovery into a second project. It can translate a brief into a reviewable shot plan, keep reference roles and acceptance criteria attached to each shot, present finished outputs, and rerun only the weak section after approval.

The closing frame tests whether identity, clothing, workshop layout, and crescent geometry survive the complete local sequence from impact to finished object.

Conclusion

A reliable MiniMax H3 local setup starts by choosing the right task family, not by downloading everything labeled H3. Confirm hardware and storage, update ComfyUI, pair one diffusion checkpoint with its matching text encoder and both VAEs, place files in the expected folders, and pass a short baseline render before adding acceleration or complex references. Version every working combination so a new node or quantization cannot erase a known-good setup. If local control is less important than planning, review, continuity, and recoverable production, start the project with Seedance Agent and use MiniMax H3 without maintaining the whole stack yourself.