LTX 2.5: Open Weights, Specs, and How to Run It

LTX 2.5 is out as of August 11, 2026, and the weights are downloadable the same day the model is announced. That is still the unusual part. The closed incumbents, Veo, Sora, Kling, Seedance, are APIs you rent, and the one top-tier rival that also publishes weights, MiniMax H3, shipped its checkpoints without the module that produces its headline resolution. LTX 2 and now LTX 2.5 are the full current model as files you can pull from Hugging Face, run on your own GPU, fine-tune on your own footage, and ship inside your own product, free of charge until your company clears $10 million in annual revenue.

This guide covers what changed in 2.5, what the license actually says, what the API costs per second, what hardware you genuinely need, how the release stacks up against MiniMax H3's, and the honest answer to whether you should run it locally at all.

LTX Video to LTX 2.5: how the family got here

The lineage matters because search results still mix the versions up. LTX Video was the original open text-to-video model from Lightricks, a 2 billion parameter release in November 2024, followed by a 13 billion parameter LTXV in May 2025. LTX-2 arrived in October 2025 as the first DiT-based audio-video foundation model that generated synchronized sound and picture in a single pass, and its full code, weights, and tooling went public in January 2026. LTX-2.3 followed in March 2026 at 22 billion parameters. LTX is Lightricks' model brand, and the weights and code still ship under the Lightricks organization on Hugging Face and GitHub. What changed with 2.5 is the framing: LTX now describes it as an open world model for physical AI, not a video generator.

LTX 2.5 is the current release, and LTX calls the family the most downloaded open world model at more than 33 million downloads. Treat that figure as the vendor's own count: LTX's public reference page still listed 18 million as of July 2026, so the number moved fast and it is not independently audited.

What is new in LTX 2.5

Four changes carry the release, and all four are aimed at production work rather than demos.

• Diffusion Fidelity Rendering. That is the name LTX uses for it on the model page and in the launch post, which also markets it as Pixel Diffusion in its summary line. One feature, two labels, so do not go looking for a second one. The model builds motion and structure in an 8x temporally compressed latent space while generating high-fidelity keyframes alongside it, allocating more compute to visually demanding moments and less elsewhere, then runs a dedicated diffusion rendering stage. A new diffusion video decoder augments the old VAE decoder. The claimed payoff is sharper faces in close-up, more legible text and signage, and fewer smears in fast motion.
• Native multishot. One generation produces several connected shots, wide to over-the-shoulder to close-up, holding character, environment, lighting, style, and voice across the cuts. This is the feature that separates a clip from a scene, and stitching shots yourself has never held identity well.
• Better prompt adherence. A custom Gemma 4 12B text encoder plus a dedicated prompt enhancer that expands short prompts into detailed instructions at near-zero compute cost. Complex multi-subject prompts survive more of what you wrote.
• Automatic duration. Send duration: null and the model predicts the clip length from the action you described instead of forcing you to pick a slot.

Alongside those: native 4K HDR output, a native EXR workflow that reads and writes inside ACES and DaVinci Wide Gamut without a lossy 8-bit round trip, and a substantially improved distilled checkpoint that carries more of the full model's quality into a smaller, faster variant. The distilled model is the one that makes local deployment realistic, so that improvement matters more than its billing suggests.

Full size

LTX 2.5 specs: resolution, frame rate, and duration

The model ships in two API variants, and they are not interchangeable. ltx-2-5-fast is the speed and cost option and reaches the highest resolutions; ltx-2-5-pro is the fidelity option and stops at 1080p. Both do text-to-video, image-to-video, and audio-to-video, in 16:9 and 9:16, with synchronized audio generated jointly with the picture.

Two footnotes worth knowing before you plan a shoot. The 20 second ceiling only exists at 720p and 1080p at 24 or 25 fps; every 1440p, 4K, and high-frame-rate combination caps at 10 seconds. And automatic duration cannot be combined with a fixed last frame on image-to-video, because a last frame fixes where the clip ends and that requires a known length.

On the open weights side, the Hugging Face repository carries 22 billion parameter transformers in both dev and distilled form, ComfyUI-ready int8 conversions of each, an NVFP4 quantized distilled variant, separate video and audio VAEs, spatial and temporal upscalers, and the Gemma 4 12B text encoder. The repo is gated, so you accept the license terms on Hugging Face before the files download.

Is LTX 2 free? The $10M license explained

Yes, for most people reading this, with one number to check. The weights ship under the LTX community license, not Apache 2.0. LTX's own licensing page states it plainly: if your company earns under $10 million in annual revenue, you can self-host the full weights, fine-tune them, and use the output commercially at no cost, with no usage caps and no seat fees. Above that line you need a commercial license negotiated with LTX.

Three details decide whether you are clear:

• It is total company revenue, not product revenue. LTX measures it across your whole organization, so a $12M agency cannot run it free on one small product line.
• Free means free of per-generation billing, not free of cost. You pay for the hardware you run it on. That is the entire economic argument: your unit cost is fixed infrastructure instead of a variable per-clip fee.
• 2.5 loosened the terms further. LTX says it reduced restrictive third-party dependencies in the license to simplify fine-tuning, redistribution, and legal review. There is also no mandatory branding on output, which is not true of every model that calls itself open.

That threshold is the whole reason this release matters to small studios. Under $10M in revenue, a frontier-class video model with native audio becomes a fixed cost you control rather than a metered API you cannot audit. Above it, LTX wants a conversation.

How much does LTX cost on the API?

If you do not want to run the weights, the LTX API bills per second of generated video. These are the published rates from LTX's pricing documentation:

So a 10 second 1080p clip on fast is about $1.30, and the same clip on pro is $1.70. One billing quirk to plan around: on prepaid accounts, automatic duration holds credits against the longest clip your resolution and frame rate allow, then releases the remainder when the job finishes. A request that would have produced 6 seconds is still declined if your balance cannot cover 20.

How to use LTX for free, and what hardware it needs

There are three routes, and only the first two are free.

1. ComfyUI. LTX shipped 2.5 with day-one ComfyUI support, which is the fastest path from nothing to a generated clip. Install the ComfyUI-LTXVideo node pack through ComfyUI Manager, open the template browser, and pick the LTX-2.5 text-to-video, image-to-video, or first-frame/last-frame template. The workflow fetches the model files it needs.
2. Python pipelines. Clone Lightricks/LTX-2 and run uv sync --frozen. The monorepo splits into ltx-core for the model and inference stack, ltx-pipelines for the generation workflows, and ltx-trainer for LoRA, IC-LoRA, and full fine-tuning. This is the route if LTX is going inside your own product.
3. LTX API or LTX Desktop. The API is the paid managed option above. LTX Desktop is a free local video editor built on the same engine, if you want the model without the node graph.

Now the hardware, because this is where the launch coverage and the documentation disagree and it will cost you an evening if you get it wrong. LTX optimized 2.5 for local inference on NVIDIA RTX GPUs and DGX Spark, and its comparison page advertises a 16GB VRAM minimum. LTX's own system requirements page for LTX-2.5 asks for an NVIDIA GPU with 32GB or more of VRAM, 32GB of system memory, 100GB of free storage, CUDA 12.7, and Python 3.12, and recommends an A100 80GB or H100. Those are different machines. The reconciliation is that quantized distilled checkpoints, the fp8 and NVFP4 variants, are what stretch down toward consumer cards, while the full 22B model wants a data center GPU. If you have a 16GB card, plan on a distilled quantized checkpoint at modest resolution, not 4K on the dev model.

Full size

Is LTX 2 better than WAN?

They solve different problems, and the honest comparison turns on what open means for each. Alibaba's Wan published downloadable weights through Wan 2.2, a 14 billion parameter mixture-of-experts model that is strong at tightly controlled image-to-video. Every Wan release since has shipped as a commercial API without publishing weights: the Wan-AI organization on Hugging Face still tops out at the Wan 2.2 series, which makes 2.2 the newest self-hostable Wan more than a year after it landed.

Against that, LTX 2.5 generates synchronized audio with the picture in a single pass, which Wan does not do natively, reaches 4K, and is materially faster per clip. LTX also keeps shipping weights on the current version rather than the previous one, which is the difference that compounds over a year. Wan 2.2 remains a reasonable choice if your workflow is silent image-to-video and you already have LoRAs trained for it. For a fuller field, our guide to open source AI video generators covers where each of these actually holds up.

LTX 2.5 vs MiniMax H3

The head-to-head that matters is not LTX against the closed APIs, it is LTX against MiniMax H3, because H3 is the only other model in the top tier claiming open weights. The two releases landed eleven days apart and the claims sound identical: downloadable checkpoints, a community license, free commercial use under a revenue cap. What is actually in the download decides it. MiniMax published two BF16 base checkpoints on August 5 but withheld H3-Regenerate-2K, the module that produces the hosted product's headline 2K output, and H3-Context-IR, the preprocessor that rewrites free-form prompts into the structured format the base model expects. A self-hosted H3 therefore tops out at 768p and takes hand-written structured prompts, and what shipped alongside the weights is deployment recipes for inference, not a training stack. LTX 2.5 publishes the whole thing it markets: the 22 billion parameter dev and distilled transformers, the diffusion video decoder, the upscalers, and a maintained trainer for LoRA and full fine-tuning, with 4K on your own hardware if the card can carry it. Our MiniMax H3 guide covers that release in full.

The licenses are cousins with different numbers. Neither is OSI open source; both are community licenses with a revenue cap on free commercial use. LTX draws the line at $10 million in total annual company revenue with no mandatory branding on output. MiniMax draws it at $20 million with attribution required. H3's cap is twice as roomy; LTX's terms are cleaner once you are under it.

Read the price row twice, because it is the one that flatters H3: hosted H3 delivers 2K for the same $1.30 per 10 second clip that LTX's fast variant charges for 1080p. That, plus a reference system that accepts up to nine images along with video and audio clips in a single request, plus first place on Artificial Analysis's video editing board, is H3's case. LTX 2.5's case is everything above the price row: 4K, longer clips, native EXR and HDR, and the only release of the two you can fine-tune with vendor tooling and run at full fidelity in your own building. If open weights are the reason you are here, LTX shipped them and MiniMax shipped a preview of them.

Where Veo 3.1 fits

Veo 3.1 keeps a column in that table because it is still the model both open releases price themselves against, and on cinematic fidelity it still wins. What it does not offer is anything in the first three rows: no weights, no self-hosting, no license threshold to check, only an account. It is also slow by LTX's measure.

LTX publishes a speed chart on its launch page showing a 10 second image-to-video clip in 6.8 seconds on its own hardware, two GB200s at steady state, and 23.7 seconds through the LTX API, against 70 seconds for Veo 3.1, 196 for Seedance 2.0, and 317 for Seedance 2.5. Read the methodology note before you quote those numbers: LTX measured its own on-prem run, took competitor timings from a third-party provider including queue time, and the models are not all rendering at the same resolution, with Veo measured on an 8 second clip rather than 10. The ranking is probably directionally right. The multiples are a vendor's.

Across all three models the tradeoff is unchanged. Open weights buy you control, fixed costs, fine-tuning on your own IP, and data that never leaves your building. They cost you a GPU, a driver stack, a node graph, and the hours it takes to learn all three. If you are building a product on video generation or you have data sovereignty requirements, that price is obviously worth paying. If you need six vertical ad variants by Thursday, it is obviously not.

The open weights are what make the second case easy to solve, though, because anyone can host them. Dream Pixel Forge runs LTX 2.5 in the browser alongside MiniMax H3, Veo 3.1, Seedance, and Grok Imagine Video on one credit balance: text-to-video and image-to-video, 720p and 1080p, 5 to 20 second clips, with synchronized audio. Same model, no GPU to buy, no ComfyUI graph to wire, no CUDA version to match. If you want to fine-tune LTX on your own footage or run it air-gapped, download the weights. If you just want the clip, generate it in a tab.

Who should actually run LTX 2.5 locally

Three groups, clearly:

• Teams building video generation into a product. Per-generation billing is the thing that kills those unit economics at scale. Owning the weights turns a variable cost into a fixed one.
• Anyone with data that cannot leave the building. Regulated industries, unreleased IP, client footage under NDA. LTX supports on-prem, edge, and fully air-gapped deployment, which no closed API can offer at any price.
• Anyone fine-tuning on their own material. LTX 2.5 ships a pretrained base checkpoint built specifically for aggressive fine-tuning, plus camera-motion LoRAs, OpenPose-driven motion, depth conditioning, and IC-LoRA video-to-video. If your work needs one character or one house style repeated forever, training beats prompting.

Everyone else is better served by running LTX 2.5 hosted, on the LTX API or anywhere else it is served. That is not a dig at open weights, it is arithmetic: a GPU capable of running the full model costs more than most people will ever spend on generated clips.

What to do this week

If you have the hardware, install the ComfyUI node pack, pull a distilled checkpoint, and test one thing specifically: whether native multishot holds a character across three cuts in a single generation. That capability is the actual news in 2.5, and it is the one no closed model on the market currently offers. We ran that exact test on release day; the three-cut render further up this page is the result. Our LTX prompt guide covers how to write for a multishot pass, which is a different job from writing a single-shot prompt.

If you do not have the hardware, run the same test on a hosted copy of the model and spend the saved evening on the prompt instead. Either way, the workflow that wastes the least money on any video model is unchanged: approve a still in the freeform generator first, then prompt motion only, as our image to video guide lays out. The MiniMax H3 guide covers the rival open-weights release, the Veo 3 prompt guide covers the model that still wins on cinematic fidelity, Seedance 2.5 covers the one going after clip length, and our roundup of AI video models in 2026 puts the whole field side by side.