--- title: Local film grew up. The closed agent behaved summary: LTX-2.3 makes the local film look like a product. GPT-5.4 turns the closed agent into a daily driver. Voxtral TTS puts a small voice on your side of the wall. The month is about finishing work, not collecting models. lede: The local film grew up, and the closed agent stopped being a stunt. date: 2026-03-31 authors: Team COEY image: /blog/what-march-2026-opened-for-production.webp image_alt: A standing portrait film slate beside a small closed black cube image_credit: COEY keywords: AI Industry News software: groundslate --- [LTX-2.3](https://ltx.io/newsroom/ltx-2-3-offers-creative-intelligence-without-the-cloud-costs) is the local video update that stopped feeling like a science project. Same family as last month's open weights. Sharper 4K. Native 9:16. Audio that does not sound like a ringtone in a hallway. A desktop app if you do not want to live in Comfy. The commercial line did not change: under ten million in revenue, you are fine. Over it, you buy a seat. We already wrote the operator's version in [LTX 2.3 Pushes Open Video Closer to Real Creative Infrastructure](/resources/blog/2026/03/22/lightricks-ltx-2-3-pushes-open-video-closer-to-real-creative-infrastructure). The month-end version is simpler. You can make a portrait ad on a box you already have, and the sound arrives with the picture. That is the production question. Not "is it 4K." Whether a junior can run a take without booking a vendor and without sending the client's face to a queue they do not control. > A desktop that renders a clip is a product. A GitHub page that renders a paper is homework. ## The film that lives on a desk LTX-2.3 is still in the twenty-billion class. It is not a toy. It is also not a cluster. Lightricks is selling the idea that a studio GPU is enough if you stop pretending every shot needs a frontier API. They are right more often than the internet wants to admit. Portrait is the unsexy win. Most of the work that pays is a story you hold in one hand. A model that treats 16:9 as the only grown-up frame is a model that does not know how ads ship. If your timeline is a feed, the model has to think in 9:16 first, not as a crop you apologise for later. The audio pass is still not a mix engineer. It is a stem you can keep. That is enough for a first assembly. It is not enough to skip a person with ears. Celebrate the magic, then hand the mix to someone who can hear when the room tone is lying. The desktop app is how this stops being a research flex. A producer who will not open a terminal can still run a take. The weights stay for the people who want a queue. Both matter. Confusing them is how you get a studio that bought a product and a researcher who cannot ship. ### Can you automate it Yes. Brief in, clip out, check the face, check the stem, refuse the take that wandered. The desktop app is for a human. The weights are for a queue. Do not confuse the two. Name the reject before you name the model. Same face across three shots. Audio that is not silence. Duration that matches the brief. If those three fail, nobody cares that the lighting was "cinematic." | Path | You keep it | You call it | | --- | --- | --- | | LTX-2.3 | Weights, local GPU | LTX API, Fal | | GPT-5.4 | Never | OpenAI | | Voxtral TTS | Weights, CC BY-NC | Mistral API | ## The closed agent that finally behaves [GPT-5.4](https://openai.com/index/introducing-gpt-5-4/) is OpenAI's "this is the one you leave on" model. Longer jobs. Fewer lost threads. Better at staying on a tool chain without writing you a speech about synergy. It is not open. It will not be. Price it like a senior freelancer you do not have to brief twice. That is a compliment. A daily driver is more valuable than a genius that needs a chaperone. The question for a marketing lead is not whether it can write a sonnet. It is whether it can take a messy brief, call the right tool, and stop when the ticket is done. **Claude Dispatch** showed up as the other closed lane: an agent that can be sent a ticket and left alone for a while. Same rule. Brilliant. A bill. Do not use it to rename a file. **NVIDIA Nemotron 3 Super** landed as the open-weight cousin of that story. Bigger than the Nano from last winter. Meant for agents that sit next to tools. If you already run Nemotron, this is the upgrade path. If you do not, do not start a new religion because a keynote had lasers. ### What a marketer can actually plug in Use GPT-5.4 when the brief is ugly and the stack is already OpenAI. Use LTX when the footage cannot leave the building. Use a small planner, not a frontier model, to decide which of those two gets the job. The planner is the unglamorous hero. FunctionGemma-class routers still exist for a reason. Spending a frontier model to pick a folder is how invoices get weird. ## A voice you can actually hold [Voxtral TTS](https://mistral.ai/news/voxtral-tts/) is the small one. About four billion parameters. An API at a price that does not require a CFO. Weights on Hugging Face if you want them, with a non-commercial licence on that path. Read that twice. The API is the commercial door. The download is for research and internal experiments unless legal says otherwise. That is still useful. A scratch voice on your metal. A production voice on a meter you can forecast. The month finally had a speech stack that was not "call a specialist house and hope." ### The test that matters Same line, three takes, same mouth shape in the picture if you are laying it under a face. If the model cannot hit that, it is a novelty voicemail. A voice that is cheap and wrong is more expensive than a voice that is priced and right. Cheap is not the same as ready. ## Closed brains stay a line item GPT-5.4. Claude. Anything with a waitlist and a brand film. LTX is the exception this month, and even LTX has a commercial ceiling. Know which side of ten million you are on before you demo it to a client. Voxtral's download is not a free commercial voice. Treat the API as the grown-up path until counsel reads the other one. ## Keep April to one film path One local film path. LTX-2.3 if you can hold it. One closed agent you already pay for, used only when the job is hard. One voice path with a licence you can explain. Do not collect models. Collect jobs that finish. If a new checkpoint does not retire a vendor or a late night, it is a bookmark. GroundSlate is where the local half sits. The clip, the stem, the brief. The rented half is a line item. This month made that easier to say out loud.