MiniMax H3 can be used through open H3-Base weights or MiniMax's hosted API. Choose open weights when you need self-hosted inference and model-level control. Choose the API when you want to generate inside an app or automated workflow without operating model servers.
The best route depends on the work you want to own. This guide shows how to start with each available option, then compare them with one repeatable test.
Open Weights or API: Which H3 Route Should You Choose?
| Choose | Best for | You manage | Main limitation |
|---|---|---|---|
| H3-Base open weights | ML teams, researchers, custom pipelines, and self-hosted deployments | Model files, compute, dependencies, inference, storage, safety, and updates | The public weights cover H3-Base at 768p, not the complete hosted H3 system |
| MiniMax H3 API | Apps, batch jobs, prototypes, and automated content workflows | API credentials, requests, task status, retries, storage, budgets, and product logic | You use MiniMax's hosted service rather than controlling the model server |
Use open weights if model files must run inside infrastructure you control. This route also fits research, custom inference code, or local pipeline experiments.
Use the MiniMax API if you want to ship a feature, automate batches, or test H3 without maintaining GPU infrastructure. You still need to build the workflow around each generation job.
If both routes fit, start with the API. It usually gives a development team the shortest path to a working H3 request. Move to open weights only when self-hosting solves a specific requirement.
How to Run H3-Base With Open Weights
MiniMax released two H3-Base checkpoint families. FL2VA handles text generation and first-frame, last-frame, or first-and-last-frame guidance. Ref2VA uses reference images, video, and optional audio to guide the result.
The public release is only the Base stage of the larger H3 system. H3-Base generates 768p audiovisual output. MiniMax's hosted Context-IR and Regenerate-2K components are not included in the current open-weight package.
Choose the checkpoint before downloading files
Pick the checkpoint from the input you actually plan to use:
| Your input | Checkpoint | Typical job |
|---|---|---|
| Text only | H3-Base FL2VA | Generate a scene from a written description |
| One opening image | H3-Base FL2VA | Animate from a fixed first frame |
| One ending image | H3-Base FL2VA | Generate toward a fixed final frame |
| Opening and ending images | H3-Base FL2VA | Build a transition between two endpoints |
| Images, video, and optional audio references | H3-Base Ref2VA | Guide identity, style, movement, camera behavior, voice, or sound |
Do not download Ref2VA just because it sounds more advanced. FL2VA is the direct choice for text and endpoint-frame jobs. Ref2VA serves a different input workflow.
Build one local workflow before adding custom components
- Read the license and model card. Confirm that your use, location, distribution plan, and organization fit the published terms.
- Choose one supported inference path. MiniMax lists SGLang, vLLM, Diffusers, and ComfyUI in the official H3 repository.
- Download the matching checkpoint and required components. Keep FL2VA and Ref2VA files clearly separated.
- Follow the official setup for your chosen runner. Pin the software versions that produce your first successful result.
- Run one simple job. Save the prompt, inputs, workflow, seed where available, and output.
- Record resource use. Note startup time, peak VRAM, generation time, storage, and failures on your own hardware.
Avoid beginning with community nodes, adapters, quantization changes, and a complex reference pack at the same time. When the first run fails, too many variables make the cause hard to find.
What local control does and does not give you
Self-hosting lets you control the serving environment, surrounding code, file handling, and compatible inference settings. It also lets you build your own preprocessing and review pipeline.
That control does not turn H3-Base into the complete hosted system. A custom preprocessor is not MiniMax's Context-IR implementation. A separate upscaler is not the official Regenerate-2K stage.
Local deployment also moves costs and responsibilities to you. Compute, storage, engineering time, monitoring, safeguards, and failed runs still matter. Privacy depends on your logs, access rules, asset storage, and third-party components.
How to Use MiniMax H3 Through the API
The API route removes model hosting, but it does not remove application work. H3 generation runs as an asynchronous task: create the task, keep its ID, check its status, and retrieve the finished output.
Start with the minimum API lifecycle
- Create a MiniMax platform account and API key. Store the key outside client-side code and public repositories.
- Choose one input pattern. Begin with text or one image-led request before adding several references.
- Create a video-generation task. Use the
MiniMax-H3model and include a non-empty prompt. - Save the returned task ID. Associate it with your user, input files, and request settings.
- Poll the task-status endpoint. Handle in-progress, successful, and failed states separately.
- Retrieve and store the result. Download the output to storage you control instead of treating the returned location as permanent.
MiniMax's official video-generation guide owns the current request fields and endpoint details. Use its current schema instead of copying an old payload from a third-party tutorial.
Add production safeguards after the first request works
Once one request completes end to end, add the surrounding controls:
- Validate prompts and media before submitting a paid task.
- Prevent duplicate submissions when a user refreshes or retries.
- Set a polling interval and a maximum wait policy.
- Separate request errors from generation failures.
- Store task IDs so interrupted sessions can recover their jobs.
- Track failed attempts and final cost per usable result.
- Apply your own access, moderation, retention, and deletion rules.
Do not scale concurrency before you can recover one failed job cleanly. A successful API call is only the start of a reliable user workflow.
Compare Both Routes With One Small H3 Test
A fair route comparison needs the same source, prompt, target, and acceptance criteria. Keep the first test short and visually simple.
Reusable test card
Source image: A licensed 16:9 image of a matte-white ceramic mug on a wooden table. The handle and printed label must be clearly visible.
Prompt:
Preserve the mug, handle, label, table, and background from the input image. Steam rises slowly from the mug as the camera makes one gentle push-in. The label remains readable. Audio contains quiet room tone and one soft ceramic clink. No music.
Acceptance criteria:
- The mug shape and handle stay stable.
- The printed label remains readable.
- Steam moves without changing the cup design.
- The shot uses one slow push-in rather than several camera moves.
- The clip includes room tone and one ceramic clink.
- No music appears.
Use the nearest shared duration and aspect ratio available in both routes. If one route cannot match a setting, record the difference instead of changing the creative goal.
Record the result, not just the render time
| Measure | What to write down |
|---|---|
| Setup time | Time from an empty project to the first accepted job |
| Generation time | Time from accepted job to completed output |
| Failures | Submission errors, runtime errors, failed tasks, and unusable files |
| Retries | Number of additional generations needed |
| Cost | Actual compute or API cost for all attempts |
| Visual result | Whether the mug, handle, label, steam, and camera move meet the criteria |
| Audio result | Whether room tone, the clink, and the no-music instruction hold |
| Follow-up work | Trimming, repair, storage, upscaling, or editing still required |
For a local run, also record hardware, runner, software versions, precision or quantization, peak VRAM, and checkpoint source. For an API run, record task states, response errors, and storage behavior.
One result cannot prove that a route is always faster or better. It can reveal which route fits your current team, infrastructure, and tolerance for failed attempts.
Fix the Most Common Setup Problems
The local workflow will not load
Check that the checkpoint matches the workflow. Then compare file names, locations, and dependency versions with the official repository. Remove third-party nodes and optimizations until the base workflow runs.
The API accepts the request, but your app loses the result
Persist the task ID as soon as MiniMax returns it. Your interface should resume status checks after a refresh or temporary connection failure. Copy completed media into your own storage.
The output ignores part of the prompt
Reduce the test to one subject, one action, one camera move, and one audio goal. Add complexity only after that smaller version works. For reference-led generation, give each source a clear job instead of asking every source to control everything.
Local and API results do not match
First compare the actual inputs, checkpoint or model label, duration, resolution, prompt, and any inference settings. The open H3-Base release and hosted service are not identical system paths, so exact output parity should not be assumed.
Which H3 Route Fits Your Work?
| You are | Start with | Why |
|---|---|---|
| An independent creator without ML infrastructure | MiniMax API | You can generate without operating a model server |
| A developer building an H3 feature | MiniMax API | The task lifecycle fits products, queues, and automated jobs |
| An ML engineer or researcher | Open weights | You can inspect and operate the released Base checkpoints directly |
| A studio with strict self-hosting requirements | Open weights | You control the deployment and surrounding data flow |
| A team unsure which route it needs | MiniMax API, then one local test | The API establishes a working baseline before you invest in infrastructure |
Keep your source assets, prompts, and acceptance criteria portable. They make it easier to compare routes without rebuilding the creative brief each time.
Frequently Asked Questions
Can I use MiniMax H3 without owning a GPU?
Yes. Use MiniMax's hosted API if you do not want to operate GPU infrastructure. You still need an application or script that creates tasks, checks their status, and stores the results.
Are H3 open weights the same as the hosted H3 service?
No. The public weights cover H3-Base FL2VA and Ref2VA. The complete system also uses hosted Context-IR and Regenerate-2K components that are not in the current open-weight package.
Can local H3-Base generate the complete official 2K result?
No. The documented local H3-Base path produces 768p output. MiniMax's official 2K path uses Regenerate-2K, which is not included in the current public release.
Should I choose FL2VA or Ref2VA?
Choose FL2VA for text generation or fixed first and last frames. Choose Ref2VA when images, video, or optional audio should guide identity, style, motion, camera behavior, voice, or sound.
Can I use H3 for commercial projects?
Possibly, but the terms depend on your access route and circumstances. Review the Community License for open weights and the current platform terms for hosted use. Check territorial, revenue, attribution, distribution, and use restrictions before release.
Choose Your MiniMax H3 Access Route
Start with the API when speed to a working result matters. Choose open weights when self-hosting or model-level access solves a requirement you can name.
MiniMax H3 is now available in DomoAI through Omni Reference.
