← Back to blog
AiAbout 7 min read

Five Video Models Landed in Azure. Fal Runs the Inference.

Published Nov 8, 2026
Five Video Models Landed in Azure. Fal Runs the Inference.

On November 5, Microsoft and fal announced that five video generation models are now available in Microsoft Foundry as a public preview, delivered as a first-party Azure service. The models come from Kuaishou, Lightricks and MiniMax. The inference comes from fal. Microsoft built neither.

That arrangement is the story. For two years the question about video models was who had the best one. The question this announcement answers is different: who gets to be the counter you buy them from. Microsoft just volunteered, and it did it by admitting it does not need to own the models to own the transaction.

The five, and what each is for

Microsoft's framing is that creative work is not one task. A team exploring several directions for a text brief, a retailer turning an approved product photo into a motion concept, a developer chaining image generation into video generation, these want different things from a model. Quality, motion, control, turnaround and cost trade off differently in each case. So the five are five positions on a map rather than five attempts at the same benchmark.

**Kling v3 Pro** is the cinematic one. Image quality plus smooth motion, native audio generation, custom elements. Its image-to-video mode lets a developer control how a specific starting frame becomes a moving shot. This is the pick when visual direction and motion performance matter more than price per second.

**LTX-2.5** ships in two tiers. Fast is for iteration and preview, Pro is for high fidelity delivery. Both do text-to-video and image-to-video, and both carry synchronized audio. The point of shipping both is that a team can draft and finish on the same platform without switching tools, which sounds minor until you have tried to match a look across two vendors.

**MiniMax H3 Max and H3 Max Turbo** are post-trained variants of H3, with the post-training aimed at instruction following, visual quality, and reasoning performance. Turbo trades some quality for faster generation and is meant for high-frequency iteration.

For scale, fal's founder Burkay Gur put it plainly: generative media applications are becoming multi-model, because different parts of the job need different things. Fal's catalogue covers more than 1,500 models across image, video, audio and 3D, plus inference, training and infrastructure. Microsoft picked five. The selection is the product.

The data residency detail

Buried in the announcement is the sentence that will actually close deals: the models are hosted in US regions, and Foundry is the first provider to offer a US-hosted Kling endpoint.

Kling comes from Kuaishou, a Chinese company. For a US enterprise with a procurement review, "we can use the model, and the inference happens in a US region on an Azure contract" is a different conversation from "we can use the model." The model is identical. The paperwork is not.

This is the part of the multi-model thesis that gets undersold. The technical argument for a broker is convenience. The commercial argument is that the broker absorbs the compliance surface. Microsoft is selling the second one.

Pricing, and a number to verify yourself

Billing runs per generated second, with resolution and audio mode moving the unit price. Microsoft says specific prices will land on the Azure pricing page over the coming weeks.

Five small glass lenses of different curvatures resting on charcoal velvet.

One flag. The official blog body and its table do not describe the billing unit the same way. Before you put this in a budget, confirm on the pricing page whether the meter reads per second or per minute, then multiply. That kind of discrepancy is common in preview announcements and it is exactly the kind of thing that turns a cheap experiment into a surprise invoice.

There is also no unified quality or latency benchmark across the five. Microsoft says as much. Selection means running your own prompts, your own source images, your own clip lengths, and your own audio requirements through each one. Preview status means no SLA and an interface that can change.

The ComfyUI trick is the useful part

A second Microsoft post, published alongside the announcement, explains how to wire Foundry image models and fal video models into ComfyUI using two custom nodes, with no API key anywhere in the chain.

Three steps.

First, deploy models in your Foundry project. Either through the portal's model catalogue or with the CLI:

az cognitiveservices account deployment create ...

The example model identifiers given are gpt-image-2.5-sunburst, FLUX.2-pro and MAI-Image-2.6 on the image side, and Fal-MiniMax-H3-Max on the video side. One naming rule matters: the image deployment name has to match the model name, because the nodes infer the model family from that string. Video deployment names can be whatever you want.

Second, give ComfyUI keyless access. The nodes use DefaultAzureCredential, which means they authenticate as whatever Entra identity the machine already has. On a laptop that is your az login session. On Azure it is the managed identity. That identity needs the Cognitive Services User role on the Foundry resource, or Foundry User. There is no key to store, rotate, or leak into a shared workflow JSON, which is the most common way ComfyUI setups get compromised.

Third, paste the project endpoint, which looks like https://xxx.services.ai.azure.com/api/projects/xxx.

The architecture is the interesting consequence. Generation happens in Foundry. ComfyUI only orchestrates. So the machine running ComfyUI does not need a GPU, and Microsoft's example is a text-to-image call feeding an image-to-video call to produce a five second cinematic shot, assembled on hardware that could be a laptop.

For anyone who has maintained a ComfyUI install with forty custom nodes and a local checkpoint zoo, this is a genuinely different shape. The workflow becomes a thin client. What you lose is the ability to work offline and the ability to swap in a local model mid-graph.

The awkward neighbor in the same news cycle

Between November 5 and 6, Microsoft's security update guide carried two CVSS 10.0 entries. One was in Entra device registration. The other was CVE-2026-85889, a missing authentication flaw in Azure AI Foundry, CWE-306, network vector, no privileges required, no user interaction.

Microsoft's note says customers do not need to take any action, because the fix landed on Microsoft's own infrastructure and there is no customer-deployed version to patch. Reported by researcher Rémy Marot, no evidence of exploitation in the wild.

Two things are worth separating here. A 10.0 score measures worst-case impact, not likelihood, and Microsoft's own exploitability assessment is low. And a cloud CVE announcement is a notification, not a task. You do not patch it. You go read the logs for unusual deployments and calls in the window before the fix.

It still belongs in the same paragraph as the launch, because both are about the same shift. When the platform runs the models and holds the keys and owns the endpoint, the security perimeter moves inside a service you cannot inspect. That is the bill for the convenience.

What to actually test

If you are evaluating this, the interesting comparison is whether a broker plus five models beats one vendor plus one model for your actual pipeline. Ranking Kling against LTX against MiniMax in the abstract tells you very little.

Run the same brief through all five. Measure the time from prompt to a usable clip, not the time to a clip. Then measure the reroll rate, because a model that needs three attempts to get motion right is three times the price regardless of the sticker. Then check whether the audio, where present, survives the edit. Sync that drifts is worse than no audio.

And keep the exit in mind. The whole appeal of a broker is that you can swap the model underneath without changing your integration. Test that by switching one and seeing what breaks. If the answer is "the prompt format," you have a broker. If the answer is "everything," you have a vendor with extra steps.

Ignite runs November 17 to 20, with a dedicated Foundry and fal session on the 19th. Pricing and general availability should follow. Until then this is a preview with no SLA, which is fine for experiments and not fine for a product launch.

Related articles