6 video AI APIs and open-weight deployment options

Magic Hour publishes this guide and includes its own product in the comparison. Treat the recommendations as editorial guidance from a vendor, and verify the linked first-party product details and your own output requirements before choosing a tool.

Video AI APIs and open-weight options at a glance

Source check: September 12, 2026. Model catalogs, endpoint schemas, pricing and licenses can change independently. Open the linked model page, API reference and license before committing a production workload.

Targeted October 19, 2026 update: the Runway catalog and routing controls below were rechecked against current developer documentation. The earlier source-check date above is retained for the broader review; current account access, prices and terms still require verification for the chosen route.

What “open-source-friendly” should mean

The useful question is how much of the stack you can inspect, replace or operate yourself. A hosted API can serve open-weight models without being open source. A downloadable checkpoint can still carry a community license with commercial or scale conditions. A visual workflow can expose an API while leaving deployment, retries and storage to your team.

• Model identity: record the exact provider, model and version used for each result.
• License: read the checkpoint and code licenses separately; do not infer commercial rights from downloadable weights.
• Execution: document queue states, cancellation, retries, webhooks and idempotency behavior.
• Data lifecycle: verify input, output and log retention before sending customer media.
• Portability: keep prompts, inputs, settings, request IDs and accepted outputs outside the provider dashboard.
• Economics: compare cost per accepted result, including failed attempts, review time and infrastructure.

A production license check before choosing open weights

LTX 2.5: check the whole entity, not one project. The binding LTX-2.x license requires a paid commercial-use agreement for entities with annual revenue of at least US$10 million, including subsidiaries and affiliates under common control. Its limited non-commercial evaluation exception is separate from production use. A smaller developer transferring a fine-tuned model or LoRA to a qualifying commercial entity does not remove the recipient’s licensing requirement. Record the entity, intended use and license version before deployment.

MiniMax H3: check geography before infrastructure. The H3 community license excludes the European Union, United Kingdom, South Korea and United States from its grant and directs interested users there to seek a separate license. It also requires prior written authorization when commercial products and services generate more than US$20 million in yearly revenue, and prominent H3 attribution in a commercial product’s interface. Downloadable checkpoints therefore do not establish permission to launch a worldwide customer product. These are documented license conditions, not a performance comparison.

1. fal.ai: broad hosted model access

fal.ai's Model APIs expose image, video, audio and multimodal models through HTTP and client libraries. Its documentation separates direct calls, blocking subscriptions, asynchronous submissions, streaming and real-time endpoints. For long video jobs, the queue documentation covers submission, status, cancellation and webhooks.

Choose fal.ai when you want to compare or ship multiple hosted models without running GPUs. Treat every model page as its own contract: inputs, output shape, price and supported controls differ by endpoint. Download results you need to retain because hosted media URLs have lifecycle rules.

2. Replicate: official, community and deployed models

Replicate's prediction API supports official models, community models and deployments. Predictions are asynchronous by default, while a wait preference can hold the connection briefly for faster jobs. The prediction object exposes status, output, errors, logs and timing data.

Choose Replicate when you want a consistent prediction lifecycle across many model publishers. Pin a version for community models when reproducibility matters. Replicate documents official models separately because they use stable owner-and-name endpoints and predictable output-based pricing; do not assume the same guarantees apply to every community model.

3. Magic Hour: managed production media workflows

Magic Hour's API exposes project-specific workflows for video, image and audio creation. The text-to-video reference documents request creation, status polling, downloadable results and credits charged. Other project endpoints can have different inputs and options.

Choose Magic Hour when the application needs several production operations behind one account and job model. Record the exact endpoint, selected model, duration, resolution, returned project ID and final credit charge. For a direct implementation comparison, use the AI video API comparison.

4. LTX 2.5: downloadable video-and-audio model weights

LTX 2.5 is a downloadable Lightricks model family for video, audio and world-simulation workflows. The official model card links the code, documentation, API and research resources. Self-hosting gives a team control over the runtime and stored media, while also making that team responsible for hardware, dependencies, serving, observability and upgrades.

Choose LTX 2.5 when reproducible local or private deployment outweighs the operational cost. Capture the checkpoint revision, inference code revision, hardware, precision, seed and settings. Read the current model and code licenses directly; “open weights” is not a substitute for a commercial-use review.

Match the checkpoint to the inference runtime. The current LTX 2.5 model card ships separate transformer, text-encoder and video/audio components. Its bf16 files support the documented ltx-pipelines route; the Comfy int8-convrot variants are ComfyUI-only and are not interchangeable with that PyTorch path. Save the selected files and compatible runtime together. A successful download is not a completed serving setup.

5. MiniMax H3: community-licensed multimodal checkpoints

MiniMax H3 provides downloadable base checkpoints for first/last-frame and multimodal-reference video generation with audio. The official repository documents supported inputs, output formats and serving routes. It also states that H3-Context-IR and the complete 2K regeneration workflow depend on hosted components or APIs that are not included with the base release.

Choose H3 when multimodal references and native audio justify a heavier deployment. Review the current community license and hosted-service dependencies before promising a fully local workflow. Distinguish results from the base checkpoints from results produced with MiniMax's hosted preprocessing or regeneration services.

6. Runway API: hosted models and routing

Runway's developer documentation exposes a managed catalog containing both Runway and third-party models. The current model list includes Gen-4.5, Seedance 2.5, Veo 3.1 and MiniMax H3. This makes Runway another hosted multi-model option, not an API limited to Runway-authored models or a downloadable-weights product.

Choose it when a supported model, editing endpoint or routing configuration fits the application. Record the model actually used, input mode, duration, resolution, task status and price. Keep web-subscription credits separate from API billing. Access to a model through a host does not grant the same deployment rights as obtaining its weights.

Choose an explicit model or a constrained router

These are two different integration decisions. An explicit model request fixes the intended model; a router delegates that choice within a configured pool. Use the former when an approved workflow must keep one reviewed model. Consider the latter when several reviewed models can satisfy the same deliverable and the application can tolerate the choice changing.

Runway’s router configuration guide documents allow lists, deny lists, an optimization preference and per-modality credit ceilings. An allow list stays closed when new models launch; the default pool and deny-list approach can admit new releases automatically. A cost preference is not the same setting as a maximum charge. Confirm both before exposing a routed feature to customers.

The generation guide uses configId plus input for routed requests; it returns the chosen model and routing metadata. An HTTP dryRun checks eligibility and estimated cost without generating an asset or billing a generation. It does not demonstrate output quality. Record the response and configuration used; do not describe a routing preview as a successful video.

Keep request fields tied to the route. The router overview distinguishes routed aspectRatio values such as 16:9 from direct-model ratio values such as 1280:720. A shared provider account does not make those payloads interchangeable. Start from the exact endpoint reference instead of renaming a field in an old example and assuming the remaining options match.

Worked choice: one approved product clip

Suppose your feature must animate an approved product image into a short landscape clip with no generated speech. This is a planning brief, not a test result. First eliminate endpoints that cannot accept the image or required settings. Next decide whether one specific model is mandatory, or whether any model in a reviewed set can meet the brief. Only then compare a direct model call with a router.

For the fixed-model route, save the endpoint, model identifier and supported settings with the approved source. For the routed route, save the configuration as well as the model returned for each result. In both cases, inspect the actual video for product changes and unintended audio. A routing decision, a completed task and an accepted creative output are three separate observations.

If the requirement is to operate the inference stack privately, neither hosted path answers it by itself. Evaluate downloadable checkpoints and their licenses, compatible runtime and complete dependencies instead. Conversely, a team needing only hosted generation should compare the host’s service terms and integration contract rather than treating a self-hosting license as the price or permission for every hosted service.

Genmo Mochi 1: check compatibility before changing an existing pipeline

Genmo Mochi 1 is a downloadable text-to-video model under Apache 2.0. It can run through Genmo’s local Python pipeline or a separately billed hosted service. Use it when an existing Mochi integration, reproducible checkpoint or customization requirement matters. Keep that deployment decision separate from Genmo’s consumer playground subscription. The official repository supplies the code, weights links, CLI, Gradio interface and Python API example.

Choose the runtime before renting a GPU. Genmo’s native single-GPU implementation documents approximately 60 GB of VRAM. The official Hugging Face model card separately lists a 22 GB Diffusers example using the bf16 variant, CPU offloading and VAE tiling. Those are different configurations, not conflicting universal minimums. Match the weights, precision and inference path; include CPU memory and transfer overhead in the deployment check.

For an API trial without operating that stack, Replicate’s genmoai/mochi-1 model page lists an approximate $0.42 per run, varying with inputs. That is Replicate’s estimate, not a measured production cost or a Genmo credit price. Use its exact Mochi API reference for integration and confirm the current charge before submitting. A Genmo web subscription does not pay this separate host.

Start with one task that fits the checkpoint. The model card describes 480p output, photorealistic strengths and possible warping under extreme motion; it also warns about animated content. Treat those as documented model boundaries and vendor observations. Save the revision, prompt, seed, frame count, inference steps and output frame rate, then inspect the downloaded clip before scaling the job. This subsection does not report an identical-prompt quality test.

Is Genmo AI free? Separate weights, playground credits and API billing

The downloadable model has an Apache 2.0 license; running it still costs hardware and engineering time. Genmo’s current pricing page lists 250 lifetime playground credits unlocked by adding a payment method, says that method is never charged for this unlock, and includes a Genmo watermark on free output. It lists 100 credits per Mochi generation. At that rate, 250 credits fund two such attempts with 50 remaining; this is allowance arithmetic, not a promise of two accepted videos.

The same page describes monthly paid credit balances resetting to the plan’s amount. Playground credits, a hosted API bill and your own GPU bill therefore cover different things. Check the current offer rather than interpreting “free open source” as unlimited hosted generation. Paid dollar prices are omitted here because they did not load reliably in this source check.

If the application needs a managed text-to-video job rather than a Mochi deployment, compare the request and result lifecycle in Magic Hour’s Text-to-Video API reference. Preserve the same task and acceptance rules, adapt the fields to the selected model, and compare completed output and total retries. This is an alternative workflow, not a claim that Magic Hour exposes Mochi or wins every model comparison.

How to run a fair API evaluation

• Define one supported text-to-video task and one image-to-video task before choosing providers.
• Use the same intent and source media, then adapt only fields required by each endpoint.
• Save request JSON, model/version, timestamps, status transitions, errors and final outputs.
• Review every attempt against motion, identity, prompt adherence, audio, artifacts and delivery requirements.
• Count failed and rejected attempts; divide total spend and engineering time by accepted outputs.
• Repeat after material model or endpoint changes instead of carrying old rankings forward.

Hosted API or self-hosted model?

Use a hosted API when time to production, elastic capacity and a documented job lifecycle matter most. Self-host when model revision control, private infrastructure or custom inference work justifies owning GPUs and operations. A hybrid architecture can prototype on a hosted platform, retain reproducible evaluation records and move only stable workloads to self-hosting.