Microsoft Foundry Brings Kling v3 Pro and Four fal Video Models to Azure

Microsoft and fal announced an integration that brings five fal-served video model variants into Microsoft Foundry as a managed Azure service. Customers can use Azure billing and Foundry’s evaluation, governance, security, and monitoring tools instead of operating separate model infrastructure.

The public preview includes Kling v3 Pro, LTX-2.5 Fast, LTX-2.5 Pro, MiniMax H3 Max, and MiniMax H3 Max Turbo. Developers can select them from the Foundry catalog alongside Microsoft’s existing language and image models.

Why fal fits the gap

Fal says its APIs cover more than 1,500 models across image, video, audio, 3D, and related generative-media workloads. The company specializes in hosting compute-intensive inference and tuning models after their initial training to improve characteristics such as speed, prompt adherence, and output quality.

Video models vary considerably in visual quality, latency, cost, audio support, and creative controls. Adding several families at once gives Foundry customers options for different workloads. Fal co-founder Burkay Gur said generative-media applications will increasingly route work across multiple models.

Five models, different trade-offs

• Kling v3 Pro: Supports text-to-video and image-to-video generation, native audio, custom elements, and directed scene controls.
• LTX-2.5 Fast: Prioritizes generation speed and supports synchronized audio for text-to-video and image-to-video workloads.
• LTX-2.5 Pro: Trades higher cost and latency for the quality-focused tier of the LTX family.
• MiniMax H3 Max: Uses fal’s post-training work to target prompt adherence, visual quality, and inference performance.
• MiniMax H3 Max Turbo: Offers the lowest published generation prices in the initial lineup and targets high-volume iteration.

The announcement provides no common quality or latency benchmark across the five variants. Teams will need to evaluate them with representative prompts, source images, clip lengths, and audio requirements.

Microsoft says the initial portfolio is hosted in the US and describes Foundry as the first provider of a US-hosted Kling 3 endpoint. Kling comes from the Chinese technology company Kuaishou. US hosting can satisfy a data-location requirement for some organizations, while a full compliance review also covers retention, subprocessors, access controls, logging, and certifications.

Ten seconds can cost 25 cents or $3

Published preview rates charge by each second of generated video. MiniMax and LTX prices vary by resolution, while Kling prices vary by audio mode. The ten-second figures below are direct equivalents of the per-second rates.

At the listed rates, ten seconds of 1080p output ranges from $0.80 on MiniMax H3 Max Turbo to $1.70 on LTX-2.5 Pro. Teams can route high-volume drafts to cheaper endpoints and use application-specific evaluations to choose a model for selected renders. Production budgets may also include storage, data transfer, moderation, retries, and downstream processing.

One control plane, model-specific code

Azure teams can manage access, billing, evaluation, and operational monitoring through Foundry, reducing separate vendor onboarding and infrastructure work. Microsoft says additional fal models will join the catalog over time.

Each model retains its own prompt behavior, controls, supported resolutions, audio modes, latency profile, and safety constraints. Switching models therefore requires model-specific testing and may require adapter code around request parameters, job status, and output handling.

Fal handles specialized model hosting and inference, while Foundry supplies the Azure-facing service layer. Google Cloud markets Veo and AWS offers Nova Reel as cloud-native video models. Microsoft’s approach adds several external model families at once, emphasizing catalog breadth under a shared enterprise control plane.

Checks before production

Public preview status warrants several checks before an application depends on these endpoints:

• API fit: Verify supported clip lengths, aspect ratios, resolutions, image inputs, audio modes, and response formats.
• Capacity: Confirm quotas, concurrency limits, rate limits, timeout behavior, support terms, and uptime commitments.
• Evaluation: Measure latency, failure rates, prompt adherence, motion quality, audio synchronization, and safety-filter behavior with representative workloads.
• Data and rights: Review processing location, retention, logging, subprocessors, output-use terms, and required compliance controls.
• Total cost: Include storage, networking, moderation, orchestration, and failed or repeated generation attempts in workload estimates.

Microsoft and fal plan to discuss the partnership at Microsoft Ignite during “Video AI at Scale with fal on Microsoft Foundry.” The Ignite session lists engineering leads from both companies.