Self-Hosted AI Video vs API: Choosing a Deployment Model

Deployment Is Now the Second Half of Model Choice

On 31 July 2026 MiniMax published the weights for H3, a 33-billion-parameter video model that produces four- to fifteen-second clips with 32 kHz stereo audio. A downloadable weight file changes the shape of a decision most production teams thought they had already settled. For two years the only question worth asking about a generative engine was how the footage held up on a grade. Now there is a second question, and it has nothing to do with craft: self-hosted AI video or a hosted API — where does the model actually run, and who is legally allowed to run it.

The two answers are not variations on a theme. Renting through an API means the vendor owns the GPUs, the safety filters, the uptime and the roadmap; you send a prompt and receive a clip. Self-hosting means downloading the weights and running inference on hardware you control, so no frame ever leaves your network. Teams already have a process for matching each engine to the job it does best, and deployment now sits alongside it as a second, colder axis of the same decision.

Most production companies will end up running both, which is fine. The useful work is deciding in advance which jobs go where, writing that rule down, and being able to answer it in the room when a client asks whether footage of their unreleased product was ever uploaded to a third party. That conversation goes badly when the answer is improvised.

Read the Licence Before You Read the Benchmark

H3 is not open source in the sense most people mean by the phrase. The weights ship under the MiniMax H3 Community License Agreement, and the open-weight release is territory-scoped: organisations in the European Union, the United Kingdom, South Korea and the United States have to apply for a formal licence before deploying the weights, while the hosted API stays globally available. The vendor's own explanation is unusually candid about why — likeness generation, copyright and content safety are regulated differently across those markets, and once weights leave the vendor's infrastructure the vendor can no longer enforce anything.

Compare that with Alibaba's Wan 2.2, which ships under Apache 2.0 and states in its repository that it claims no rights over the content users generate. Apache 2.0 also grants an irrevocable patent licence and, under Section 4, requires anyone redistributing a derivative work to pass on the licence, mark modified files and preserve the original copyright and attribution notices. That last clause matters more than it sounds. Fine-tune a model on a client's brand, hand over the resulting checkpoint, and you have redistributed a derivative work.

So the licence review is four questions, asked before anyone benchmarks anything. Does this licence permit commercial use at our revenue and in our territory? Does it cover the outputs, or only the weights themselves? Does it survive being handed to a client at the end of a retainer? And what happens if the terms are rewritten mid-campaign? Weight licences sit on top of, not instead of, the clearance work behind keeping generated footage commercially safe, and both answers have to be right at the same time.

What Self-Hosted AI Video Actually Buys You

Three things justify running the model yourself, and per-clip price is not one of them. The first is containment. Unreleased packaging, pre-embargo talent plates, unannounced product geometry and anything covered by a strict non-disclosure agreement can be generated without a single frame crossing a vendor boundary. For regulated categories, and for clients whose legal team actually reads the data-processing addendum, that is the entire argument.

The second is adaptation. Prompt engineering plateaus quickly on brand-specific form factors — a particular bottle silhouette, a recurring set, a house lighting style the model keeps softening back towards its own average. A small fine-tune on a couple of hundred approved frames does what a longer prompt cannot, and only published weights make it possible. The third is stability: a local checkpoint is frozen, so the model that approved the animatic is the same model that renders the master.

That last point is badly underrated. Hosted models are updated silently, and a version bump landing between client approval and final delivery can quietly shift skin tone, motion cadence or grain structure. On a campaign shot across six weeks, that is a continuity problem wearing the costume of an upgrade. Owning the weights means owning the version, which is the same discipline as locking a LUT before the grade starts rather than after.

What It Costs You Before It Saves You

Self-hosting trades a predictable invoice for an engineering project. A top-tier GPU rented by the hour looks cheap in isolation and is expensive when idle, and production demand is bursty by nature — three frantic days, then two quiet weeks. Owned hardware inverts the problem: the capital is spent whether or not anything renders. Neither shape matches how an agency bills, which is why the maths has to be run on real monthly volume rather than on the peak week that prompted the conversation.

Then there is the part nobody budgets. Someone has to keep the cluster alive, update drivers, manage checkpoints and be reachable the night before a delivery. That is a headcount line, not a tools line. You also inherit the safety filtering the vendor used to run, because the model does not arrive with guardrails attached, so content moderation quietly becomes an internal process with an owner and a review step. The honest comparison runs against your real cost per usable clip, not against a feature list.

A workable rule of thumb: if generative video is a recurring, high-volume line in the business with reasonably predictable monthly demand, self-hosting starts to earn its keep. If the work is project-driven and spiky, the API almost always wins, and the engineering time is better spent on the review and approval pipeline, which is where the real bottleneck sits in nearly every team that has measured it.

Compliance Does Not Travel With the Weights

There is a tempting misreading of self-hosting: that keeping everything in-house makes the disclosure question quietly go away. It does the opposite. Obligations to label synthetic content attach to whoever puts the advertising in market, not to whoever owns the GPU. Running the model in-house does not shrink the labelling obligations that already apply to AI advertising; it transfers the whole of it onto your side of the contract, along with the evidence you would need to defend a decision eighteen months later.

Provenance makes this concrete. Under the C2PA specification, a manifest is produced by a claim generator, and from version 2.0 onwards that claim generator must also be the signer — the holder of the private key, identified by the subject of the certificate. When a vendor generates the clip, the vendor's software signs it and the trust chain points at them. When you self-host, that signature is yours to issue, which means obtaining a certificate, managing the key and deciding what your assertions actually assert.

None of this is an argument against self-hosting. It is an argument for staffing it properly. The teams that get this wrong are the ones who treat a weight download as a procurement decision, then discover six weeks later that nobody owns disclosure, nobody owns signing, and nobody kept a record of which checkpoint produced the master that shipped.

A Routing Rule You Can Write Down Today

Make the API the default, because it usually is the right answer. Exploration, animatics, mood tests, high-volume social variants and anything built from material the client has already made public should run through a hosted model, where the cost is variable and the maintenance belongs to somebody else. Reserve self-hosting for the narrow set of jobs where one of four conditions genuinely holds, and require that all four be checked rather than assumed by whoever is keenest on the idea.

The conditions are simple to state. The material is confidential enough that a vendor boundary is unacceptable. The output needs a fine-tune no API will expose. Monthly volume is high and stable enough to clear the break-even honestly. And the weight licence clearly permits commercial use at your revenue, in your territory, for your client list. Miss the fourth and the first three stop mattering, because the cheapest render in the world is worthless if it cannot legally ship.

In practice this produces a hybrid pattern rather than a winner: explore on the API, lock the look, then render the confidential finals locally. Review the rule quarterly rather than annually, because licence terms are moving faster than the benchmarks right now — territory scopes widened, community licences rewritten, pricing repositioned within weeks of a launch. A deployment policy written in January will be describing a market that no longer exists by summer, and the cost of that drift always lands on the one campaign where it matters most.