Nvidia's Local AI Week Continues: LTX-2.5 Video Model and Nemotron 3.5 Lightning Land the Same Day

Since Nvidia opened the week with a cheaper-routing model and an open-sourced storage stack, the company has kept shipping. On Tuesday, August 11, 2026, two more releases landed under Nvidia's banner: LTX's video generation model LTX-2.5 and Nvidia's own Nemotron 3.5 Lightning agent model. Both are open weights. Both are built to run on hardware people already own.

LTX, the generative media company spun out of Israeli app maker Lightricks, released LTX-2.5 as the newest version of its open-weights video and "world" model. It launched with native integration into ComfyUI, the node-based workflow tool that has become the standard prototyping environment for open generative media, through what both companies describe as a day-one launch partnership.

The headline number is speed. LTX says a 10-second video clip generates in 6.8 seconds running on-premises on two Nvidia GB200 chips. Through LTX's managed API, the same clip takes 23.7 seconds. LTX's published benchmark lists closed competitors Omni Flash, Grok 1.5, and Veo 3.1 at 52 to 70 seconds, with Seedance 2.0, FLUX 3, Seedance 2.5, and Kling 3.0 Pro running from 196 to 398 seconds.

VentureBeat pushed back on one of LTX's other claims. The company said LTX-2.5 costs roughly one-eighth as much and renders seven times faster than comparable models. VentureBeat checked that against published pricing and found the cost multiple "doesn't survive contact with the models that publish pricing," noting LTX-2.5's Fast tier runs $0.09 per second of 720p video with audio, putting a 10-second clip at $0.90, which VentureBeat called genuinely cheap but not the eightfold gap LTX advertised.

The model is free for organizations under $10 million in annual recurring revenue, with larger companies negotiating a commercial license. LTX says its model family has passed 33 million downloads on Hugging Face, which it calls the most-used open world model line on the market.

Technically, LTX-2.5 rebuilds most of the generation pipeline rather than adding features to the prior version. Changes include a new diffusion video decoder aimed at reducing artifacts in high-motion footage, native multishot generation that holds character and scene consistency across cuts in a single output, a custom Gemma 4 language backbone for handling complex prompts, and a checkpoint pretrained for physical AI and robotics work. The model is optimized for local inference on Nvidia RTX GPUs, letting studios keep footage and intellectual property off outside servers.

Nvidia released Nemotron 3.5 Lightning the same day, a new entry in its existing Nemotron 3 family aimed at high-volume agentic tasks. Nvidia generative AI VP Kari Briski described it during a briefing as packing "the knowledge of Ultra into the size of Nano." It ships alongside NeMo Switchyard, a new routing library meant to direct specialized tasks to the right model, and is built to run locally for added security.

The timing appears deliberate. Nvidia and even historically closed-model companies like OpenAI have leaned harder into open-weight releases following recent security incidents, while Chinese startup DeepSeek has been raising prices on its open models, shifting the competitive landscape toward Nvidia's favor. Meta released its own open agentic model, Muse Glimmer, for local use on Macs and PCs just days earlier, and Meta CEO Mark Zuckerberg published an essay Monday, August 10, arguing that open-weight models will spread AI's benefits more broadly.

Both releases plug into a broader argument playing out across the industry over open versus closed AI models. Proponents, including Zuckerberg, argue open weights let outside researchers and cybersecurity experts inspect how a model works, which builds trust and lets smaller companies compete without paying cloud fees for every generation. Critics of unrestricted open-weight releases counter that publishing full model weights makes it easier for bad actors to strip out safety guardrails, a concern that reportedly fed into the Trump administration's consideration of restrictions on foreign AI models this past July. Industry figures pushed back on that idea, arguing a ban could hurt cybersecurity rather than help it.

What's unresolved is whether LTX's benchmark numbers hold up under independent testing outside the company's own published comparisons, and whether Nvidia's local-execution pitch for Nemotron 3.5 Lightning actually reduces security exposure compared to cloud-hosted alternatives, or simply shifts where the risk sits. Nvidia's local AI series is scheduled to continue through the rest of August.