LTX-2 Now Production-Ready on SGLang x GMI Cloud

GMI Cloud and SGLang have jointly integrated LTX-2, a 19B-parameter open-source video generation model, into the SGLang inference framework - production-validated and available now. For teams already running SGLang for LLM workloads, this means text, image, and video generation through a single API, without managing a separate serving stack for each modality.

What you'll learn in this article:

●       Why integrating a 19B-parameter video model into a production serving framework required more than a straightforward port

●       How GMI validated multi-GPU parallelism, CPU offloading, and pixel-level output fidelity against reference outputs

●       What LTX-2 supports: text-to-video, image-to-video, and joint video-plus-audio generation through the SGLang API

●       How this integration eliminates the need for separate inference pipelines for language and video workloads

●       How to deploy LTX-2 today on GMI Cloud infrastructure via dedicated endpoint or closed-source API

Overview

GMI and SGLang have jointly integrated LTX-2, a 19B-parameter open-source model for text-to-video, image-to-video, and joint video-plus-audio generation, into the SGLang inference framework, production-validated and available now.

Teams Can Now Use LTX-2!

What does LTX-2 in SGLang mean for teams building multimodal products?

LTX-2 in SGLang consolidates text and video inference into a single API and a single operational stack. Teams no longer need to run a separate pipeline for video generation alongside their LLM infrastructure - one framework handles both, reducing operational overhead for multimodal product development.

SGLang has been the go-to high-performance serving framework for LLM inference. It's fast, it's well-maintained, and a lot of teams have already built their inference stack around it. 

Until now, if you also needed video generation, you were running a second stack alongside it. That's no longer the case.

LTX-2 in SGLang means text and video inference through a single API and a single operational model to maintain. For teams building multimodal products, or just trying to avoid the overhead of managing separate pipelines for every new modality, this is a real consolidation.

The other thing worth noting: LTX-2 is one of the strongest open-source video generation models available. Accessible through the same SGLang API you'd use for any LLM workload, the barrier to experimenting with and eventually shipping it drops considerably.

The Technicals

How did GMI Cloud validate LTX-2 for production use in SGLang?

GMI's engineering team validated multi-GPU parallelism on 8-GPU clusters, implemented CPU offloading for memory-efficient deployment, and tuned throughput for sustained API load. Output quality was verified against reference outputs at the pixel level - the SGLang implementation produces results that match what the model is designed to generate.

Getting a 19B-parameter video model into a production serving framework is not a straightforward port. The research environment and the production environment have almost nothing in common in terms of what they demand from the model.

GMI's engineering team worked directly with SGLang to close that gap. That meant validating multi-GPU parallelism on 8-GPU clusters, implementing CPU offloading for memory-efficient deployment, and tuning throughput for the kind of sustained API load that real workloads generate. 

It also meant running the SGLang implementation against the reference outputs at the pixel level because quality drift under a serving framework is a real failure mode that's easy to miss and hard to debug later.

The integration passed. The SGLang implementation of LTX-2 matches the reference outputs at the pixel level. What you get through the API is what the model is supposed to produce.

Here's our internal benchmarks:

Try it now

LTX-2 is available today through SGLang on GMI's GPU infrastructure.

If you want to run it yourself, get started on GMI Cloud with a dedicated endpoint or try our closed-source versions. If you're evaluating it for a specific workload — video API, multimodal product, high-volume generation pipeline — talk to the GMI engineering team about deployment architecture and what the infrastructure actually looks like at scale.

Shout-out and Special Thanks

We’d like to give a special shout-out to FlamingoPg from Sglang Community for his outstanding contributions to the SGLang LTX-2 support. He contributed thousands of lines of code, including major work on the LTX-2 VAE and the integration of various audio-video processing components. His efforts significantly strengthened the stability and completeness of the LTX-2 pipeline within SGLang. We truly appreciate his dedication and impact to the community.

Frequently Asked Questions

1. What is LTX-2 and what can it generate?
  LTX-2 is a 19B-parameter open-source model developed by LTX. It supports text-to-video, image-to-video, and joint video-plus-audio generation. It is one of the strongest open-source video generation models currently available and is now accessible through the SGLang API.

2. What did the SGLang integration require technically?
  Integrating a 19B-parameter video model into a production serving framework required validating multi-GPU parallelism on 8-GPU clusters, implementing CPU offloading for memory-efficient deployment, tuning throughput for real API load, and verifying pixel-level output fidelity against reference outputs to catch quality drift early.

3. How does this integration benefit teams already using SGLang?
  Teams running SGLang for LLM inference can now add video generation to the same stack without deploying or maintaining a separate pipeline. LTX-2 is accessible through the same API used for any LLM workload, lowering the barrier to experimenting with and eventually shipping video generation in production.

4. How can I deploy LTX-2 on GMI Cloud today?
  LTX-2 is available through SGLang on GMI's GPU infrastructure. You can get started on GMI Cloud with a dedicated endpoint or access closed-source versions through the GMI Console model library. For high-volume or specific workload deployments, the GMI engineering team can advise on deployment architecture.

5. Who contributed to the SGLang LTX-2 integration?
  The integration was built jointly by GMI Cloud and the SGLang team, with significant open-source contributions from FlamingoPg from the SGLang community, who contributed thousands of lines of code covering the LTX-2 VAE and audio-video processing components.