Open-Source AI Video Generators in 2026: What You Can Run

An open-source AI video generator is a video model whose weights you can download and run on your own GPU, instead of renting it through someone else’s app. As of October 17, 2026, the shortlist worth your time is Wan 2.2, HunyuanVideo 1.5, LTX-2.5, MiniMax H3, Kandinsky 6.0 Video and a handful of newer audio-video releases. The big change this year is sound: several open models now generate synchronized audio with the picture. The hard parts haven’t changed. Clips are short, the best ones need serious hardware, and “open” licenses come with real conditions.

Quick answer: the best open-source AI video generators right now

• Best all-round, cleanest license: Wan 2.2 (Apache 2.0, about 5-second clips at 480p or 720p, no audio).
• Best with sound you can use commercially in the US: LTX-2.5 (free under $10M yearly revenue) or Kandinsky 6.0 Video (MIT, 5-second clips with 44 kHz audio).
• Best for 12 to 16 GB cards: HunyuanVideo 1.5 (8.3B parameters, about 14 GB with offloading, no audio).
• Most capable, with a catch: MiniMax H3. Its license doesn’t cover the United States, EU, UK or South Korea.
• Not open at all: Wan 2.5 through Wan 3.0, Kling, Seedance, Veo and Runway are hosted or API-only.

Every spec below links to or names its source with the date we checked it. Model pages change fast, so re-check the model card before you download anything. Disclosure: we make CinemaDrop, a hosted AI filmmaking studio. It doesn’t host any of the open models in this guide, and we say plainly where a hosted studio is or isn’t the better choice.

Open-source AI video generators compared (October 2026)

This table covers models with downloadable weights and a published license. “Clip length” is what one generation produces on the official pipeline; “GPU” is the lowest figure the developer publishes unless noted. All entries checked October 17, 2026, unless marked otherwise.

Sources: the official Hugging Face model cards and license files for each model (checked Oct 17, 2026); Lightricks’ LTX blog for the 32 GB minimum (May 13, 2026); Quantslant’s README review for Agnes (Oct 12, 2026); the Prism license mismatch was logged in our research notes on Oct 6, 2026. Older open models such as CogVideoX 1.5 (Nov 2024) and Mochi 1 (Oct 2024) still work but haven’t had a new release since 2024 (their GitHub pages, checked Oct 17).

What counts as open source (and why the license matters)

Most “open-source” video models are really open-weight: you can download the checkpoint, but the training data stays private and the license can limit how you use it. For a filmmaker or a business, the license matters more than the parameter count. Here’s how the current models split.

Truly permissive: Apache 2.0 and MIT

Wan 2.2, MAGI-2 Preview and Agnes-Video-2.0-MoE use Apache 2.0, and Kandinsky 6.0 Video uses MIT. You can use them commercially with no revenue cap or region limit (see the Wan 2.2 model card and Kandinsky 6.0 model card, checked Oct 17). Check bundled parts, though: Agnes ships with a Gemma text encoder that has its own terms (Quantslant, Oct 12).

Community licenses with conditions

• MiniMax H3 doesn’t cover the United States, the EU, the UK or South Korea. The MiniMax H3 Community License (dated Aug 2, 2026) defines those as “Excluded Territories” and says use there isn’t authorized without a separate license from MiniMax. It also requires commercial products to display “MiniMax H3” and asks companies above $20 million in yearly revenue to get written permission.
• HunyuanVideo 1.5 uses the Tencent Hunyuan Community License, which “does not apply in the European Union, United Kingdom and South Korea.” Services with more than 100 million monthly active users need a separate license.
• LTX-2.5 is free for commercial and production use for organizations under $10 million in yearly revenue; above that you need a paid agreement with Lightricks (LTX-2.5 model card, checked Oct 17).

Both the H3 and Hunyuan licenses also forbid using the model or its outputs to improve other AI models. None of this is legal advice. Read the license file itself before you ship client work.

Open models that generate sound

Until this year, synchronized dialogue and sound effects were mostly a closed-model feature. Now there are several open options. Here’s what each one actually does.

LTX-2.5 (Lightricks)

LTX-2.5 is a 22-billion-parameter model with open weights, released August 11, 2026, that generates video and audio in one pass. Its headline feature is native multi-shot generation, so one prompt can produce connected shots, and the model card’s example pipeline upscales 2x to about 1088x1920 (LTX-2.5 model card, checked Oct 17). It runs in ComfyUI. Lightricks says it needs an NVIDIA GPU with at least 32 GB of memory (LTX blog, May 13, 2026). One tester ran 14 LTX LoRAs on a 12 GB card (Loop Forge, Oct 9). On blind-preference leaderboards it sits well behind the closed leaders, around 21st to 23rd on Artificial Analysis’ text-to-video board (as logged in our research notes, Oct 16, 2026).

Kandinsky 6.0 Video (Kandinsky Lab)

Open-sourced October 6, 2026, under MIT. Lite (3B) and Pro (29B) generate 5-second clips with synchronized 44 kHz audio, including lip sync, from text or from an image. A separate super-resolution model upscales to 1920x1080 (Kandinsky 6.0 model card, checked Oct 17). It runs through Diffusers, ComfyUI and vLLM-Omni. The team offers a 16 GB memory preset using block offloading (a company figure we haven’t tested), and a Pro clip in Full HD takes about 402 seconds on an H100 (AICoder summary of the release, Oct 6). It isn’t on the Artificial Analysis leaderboards yet.

MiniMax H3 (MiniMax)

H3 is the most capable open video model on paper. The released part, H3-Base, is a 33B dense model that makes 4- to 15-second clips at 768p and 24 fps with 32 kHz stereo audio, and ComfyUI templates exist for it. Two pieces aren’t open: the prompt-processing step (H3-Context-IR) and the 2K regeneration step are API-only (MiniMax H3 model card, checked Oct 17). The hosted H3 Max ranks first on Artificial Analysis’ image-to-video board (as logged Oct 16), but that’s the full hosted system, not what you run locally. And again: if you’re in the US, EU, UK or South Korea, the license doesn’t cover you.

MAGI-2 Preview and Agnes-Video-2.0-MoE

Two big Apache 2.0 audio-video models are also downloadable. MAGI-2 Preview from Sand.ai is a 114B mixture-of-experts model that activates 6B parameters per token; it makes 10-second clips refined to 1080p, the full download is about 307 GB, and the card lists eight NVIDIA Hopper GPUs as the requirement (MAGI-2 model card, checked Oct 17). Agnes-Video-2.0-MoE arrived on Hugging Face on October 11 with 28.7B total and 15.3B active parameters and about 57.5 GB of weights. Its README doesn’t list GPU requirements, maximum resolution or clip length (Quantslant, Oct 12). Treat both as data-center models for now.

Prism preview (Tencent Hunyuan and Fudan University)

Prism, released as a preview on October 6, targets native high-resolution video with sound. Inference at 720p needs a single 80 GB GPU, and 1080p or 2K needs four or more; training clips top out at 289 frames, around 12 seconds (Prism model card). There’s no ComfyUI node yet. Its license is listed as MIT on Hugging Face but “Other” on GitHub, so don’t assume MIT.

Silent open video models still worth running

Wan 2.2 (Alibaba)

Released in July 2025 under Apache 2.0, Wan 2.2 is still the newest Wan you can download. The A14B models use two 14B experts (27B total, 14B active per step) and make 5-second clips at 480p or 720p; the single-GPU command in the official README needs at least 80 GB of VRAM. The smaller TI2V-5B model does text- and image-to-video at 720p and 24 fps on an RTX 4090, in under nine minutes for a 5-second clip (Wan 2.2 model card, checked Oct 17). It also has a large community LoRA ecosystem.

Wan 2.5, 2.6, 2.7 and 3.0 are API-only. On October 13, neither the Wan-AI organization on Hugging Face nor Wan-Video on GitHub had any Wan 3.0 weights, code or license (iimagined.ai, Oct 13), even though Wan 3.0 tops Artificial Analysis’ text-to-video-with-audio board (as logged Oct 16).

HunyuanVideo 1.5 (Tencent)

Released November 20, 2025, HunyuanVideo 1.5 is still Tencent’s newest open HunyuanVideo; there is no official open “HunyuanVideo 2.0.” It’s an 8.3B model for text- and image-to-video at 480p and 720p, with an upscaler to 1080p, and it needs a minimum of 14 GB with offloading. A step-distilled 480p image-to-video model renders in about 75 seconds on an RTX 4090, per Tencent, and the community Wan2GP build runs it in as little as 6 GB (HunyuanVideo 1.5 model card, checked Oct 17). It makes no sound; Tencent publishes a separate HunyuanVideo-Foley model for sound effects.

CogVideoX and Mochi 1

Both still show up in “best of” lists, but neither has had a new release since 2024. CogVideoX 1.5 dates from November 2024, and Mochi 1 (October 2024) is a 480p model that needs about 60 GB of GPU memory (their GitHub pages). They’re fine for learning, but the models above have passed them.

What’s new in open-source AI video since October 6

Open video moves weekly. Here’s everything we logged between October 6 and October 17, 2026.

• Oct 6: Kandinsky 6.0 Video (MIT) and the Prism preview both shipped with audio-video generation (see above).
• Oct 7: SGF+, an Apache 2.0 project built on Wan 2.1 1.3B, streams one continuous video for up to 24 hours at 480p-class quality, with no ComfyUI node yet (ArtRealmAI, Oct 8).
• Oct 7: TaoMate-H3 released LoRA weights and inference code that add streaming audio-video generation on top of MiniMax H3, at three denoising steps per segment (ChatPicture, Oct 7). Because it builds on H3, H3’s license terms still apply.
• Oct 11: Agnes-Video-2.0-MoE weights landed on Hugging Face under Apache 2.0 (Hugging Face; Quantslant, Oct 12).
• Oct 14: ATAS, an MIT-licensed research framework, turns a script into shot prompts with scene and event memory to keep long videos consistent. Its authors tested it on MiniMax H3, Wan 2.7, HappyHorse 1.0 and Wan 3.0 (mlllm.io, Oct 15).
• Coming Oct 20: Nuva Lab says it will release open weights for FastH3-Omni, a model built on MiniMax H3 (Nuva Lab blog, Oct 14). Not released at the time of writing.
• Still closed: no Wan 3.0 weights, and no new HunyuanVideo, CogVideoX or Mochi version.

Can I run it? GPU and setup reality

“Free” open-source video still costs GPU time, setup time and electricity. Here’s a realistic map by VRAM, using developer figures unless noted.

Two tools do most of the work. ComfyUI runs Wan 2.2, HunyuanVideo 1.5 and LTX-2.5, and has templates for MiniMax H3 and support for Kandinsky 6.0. Diffusers is the Python route if you’d rather script your pipeline. Expect minutes per clip on consumer cards, not seconds. Lightricks’ headline of a 10-second 720p clip in under 7 seconds was measured on a pair of NVIDIA GB200 superchips (LTX blog, May 13, 2026), not on a desktop GPU.

Open vs closed AI video quality in October 2026

The honest answer: open models are closer than ever on features, but the top of the quality leaderboards is still closed or API-only. On Artificial Analysis’ text-to-video-with-audio board (as logged in our research notes on Oct 16, 2026), the top three were Wan 3.0 (API-only), Utopai X (closed) and Seedance 2.5 (closed). On the main text-to-video board, LTX-2.5, the best-known open model with sound, ranked around 21st to 23rd. Kandinsky 6.0 wasn’t listed yet.

Length is the other gap. Open models generate about 5 seconds (Wan 2.2, HunyuanVideo 1.5, Kandinsky 6.0) to 15 seconds (MiniMax H3) per clip. Some closed models now go much longer: Seedance 2.5 generates 4 to 30 seconds per clip (see our Seedance 2.5 guide), and Kling 4.0 is another closed option to compare (see our Kling 4.0 guide). If you need finished, polished hero shots on a deadline, closed models still win. If you need privacy, custom LoRAs or unlimited drafts on hardware you already own, open models are now good enough to be useful.

A 5-second clip isn’t a scene: the continuity gap

Most open-source AI video guides stop at “here’s a clip.” But a film is dozens of shots that have to match: the same face, the same jacket, the same room, the same voice. That’s where most local workflows break down, because each generation starts from scratch.

What helps, whichever model you use:

• Plan shots before you render. Write a shot list with sizes, angles and one camera move per clip. Our guide to types of camera shots has prompt phrases that open and closed models both understand.
• Lock your character with references. Use the same reference images every time, and train a LoRA if your model supports it (Wan 2.2, LTX-2.5 and HunyuanVideo 1.5 all do). See our guide to consistent AI characters.
• Chain clips in your local pipeline: export the last frame of one clip and use it as the start image for the next. It keeps continuity but slowly drifts, so re-anchor to your reference images every few shots.
• Use multi-shot generation where it exists. LTX-2.5 can generate connected shots in one pass, and research frameworks such as ATAS and SGF+ target long-video consistency directly.
• Keep audio consistent. Open models that make sound generate it per clip, so dialogue tone and room sound can change between shots. Plan to mix in post.

A practical hybrid: experiment and build looks locally, where drafts are free once you own the GPU, then build the actual story somewhere that keeps characters and shots organized. Our hybrid AI filmmaking workflow guide walks through that split in detail.

When a hosted studio is the better deal

Run open models yourself if you have the GPU and you need privacy, fine-tuning or unlimited experiments. A hosted studio is usually the better deal when:

• you don’t own a 24 GB or larger GPU, or your time is worth more than setup and render queues;
• you need longer shots or the top-ranked closed models, which aren’t downloadable;
• the open model you want isn’t licensed where you work (MiniMax H3 in the US, EU, UK or South Korea, or HunyuanVideo 1.5 in the EU, UK or South Korea);
• you’re making a story, not a single clip, and need characters and shots to stay organized across a whole project.

That last case is what CinemaDrop is built for. Paste a script into Script to Storyboard and it’s split into scenes and shots. Turn characters, locations and props into Elements with up to four reference images, and tag them with @ in shot prompts so the same character carries through. Then animate shots with hosted models such as Seedance 2.5, Kling 3.0, Veo 3.1 or Wan 2.7, and export all your storyboard assets as a ZIP to finish in your editor. CinemaDrop doesn’t run the open models in this guide. The free Lite plan includes 50 one-time credits for scripts, images and storyboards; video generation starts on the paid Basic plan, and Seedance 2.5 needs the Plus plan. If you’re new to the whole process, start with our beginner’s guide to AI video.

Open-source AI video generator FAQ

Is there a free open-source AI video generator with sound?

Yes. Kandinsky 6.0 Video (MIT) and MAGI-2 Preview and Agnes-Video-2.0-MoE (both Apache 2.0) generate synchronized audio with no license fee or revenue cap. LTX-2.5 is free for organizations under $10 million in yearly revenue. MiniMax H3 also makes sound, but its license excludes the US, EU, UK and South Korea. “Free” means no license fee: you still pay for the GPU.

Can I use open-source video models commercially?

It depends on the license. Wan 2.2, Kandinsky 6.0, MAGI-2 and Agnes-Video-2.0-MoE allow commercial use without revenue or region limits. LTX-2.5 is free under $10 million in yearly revenue. HunyuanVideo 1.5 isn’t licensed in the EU, UK or South Korea, and MiniMax H3 isn’t licensed in the US, EU, UK or South Korea. Always read the license file for the exact version you download.

What GPU do I need to run an open-source AI video model?

For a comfortable start, 16 to 24 GB: HunyuanVideo 1.5 needs 14 GB with offloading, Wan 2.2’s 5B model runs on an RTX 4090, and Kandinsky 6.0 has a 16 GB preset. LTX-2.5 officially needs 32 GB. Wan 2.2 A14B, MiniMax H3 at full quality, Prism and MAGI-2 are 80 GB-class or multi-GPU models.

Which open-source model is best for image-to-video?

MiniMax H3 is the strongest where its license applies; its hosted H3 Max version ranks first on Artificial Analysis’ image-to-video board (as logged Oct 16). Elsewhere, Wan 2.2’s I2V-A14B, HunyuanVideo 1.5’s I2V models and LTX-2.5 are the main choices.

Can open-source models make long videos?

Not in one clip. Most generate about 5 seconds; MiniMax H3 reaches 15. Longer videos come from chaining clips, LTX-2.5’s multi-shot mode, or research methods like SGF+, which streams a single continuous 480p-class video for hours.

Is Wan 3.0 open source?

No. Wan 3.0 is available only through Alibaba’s apps and API. Wan 2.2 is the newest Wan with open weights (as of Oct 13, 2026).

What’s the latest open-source HunyuanVideo?

HunyuanVideo 1.5, released November 20, 2025. Tencent hasn’t released an official open “HunyuanVideo 2.0” as of October 17, 2026.

The bottom line

Open-source AI video is finally useful for real work: Wan 2.2 for clean licensing, HunyuanVideo 1.5 for mid-range GPUs, LTX-2.5 and Kandinsky 6.0 for sound, and MiniMax H3 if your country is covered. Pick by license first, then by GPU, then by whether you need audio. When you need longer shots, top-ranked models or a whole story that holds together, pair your local experiments with a hosted studio rather than forcing one tool to do everything. We’ll update this guide as new open weights ship.