Correction, 2026-10-02: this page left out MiniMax H3, whose weights MiniMax published on Hugging Face on 2026-07-28. That made two claims wrong: LTX-2.5 is no longer the most downloaded open video model on the list (H3 had about 3.57 million downloads in the past 30 days, LTX-2.5 about 1.6 million, both checked 2026-10-02), and it is not the only open release here that generates synchronized audio (H3's open checkpoints output video and stereo audio). H3 is now in the list, the TL;DR and the FAQ, with its main catch: its community licence does not cover the United States, the United Kingdom, the European Union or South Korea. LTX-2.5 stays the pick, and the reason is now stated.
Correction, 2026-10-01: we said we host Wan 2.2, next to a description of its text-to-video and image-to-video models, without saying that on 8frame Wan 2.2 runs only from a start image or an input video. For a text prompt on 8frame, the Wan model is Wan 3.0 (24 credits for 5 seconds at 480p). The open-weights section is unchanged: run locally, Wan 2.2 does text-to-video. We also named LTX-2.5's license the LTX-2 Community License; its model card and repository now license it under the LTX-2.x Community License Agreement (dated 2026-08-11, same $10M revenue threshold). And we dated Mochi 1's preview weights to 2025; they were published on Hugging Face on 2024-10-22.
Correction, 2026-09-30: we called LTX-2.5's download count seven figures (it was 912,729), gave a used-GPU payback figure we had not sourced, and still listed Sora 2 as available; it was retired on 2026-09-24. Fixed.
The best open source AI video generator in 2026 is LTX-2.5 from Lightricks: current-generation weights, synchronized video and audio from text, image, or video input, and a community license that allows commercial and production use at no cost below $10M in revenue. It is not the most downloaded: MiniMax H3 had about 3.57 million downloads in the past 30 days against about 1.6 million for LTX-2.5 (Hugging Face, checked 2026-10-02). LTX-2.5 stays the pick because its licence reaches readers H3's does not: the MiniMax H3 Community License excludes the United States, the United Kingdom, the European Union and South Korea, so a reader there cannot use H3 under it.
The rest of the field is smaller than the marketing suggests. Most of the models people call open in 2026 are not: Veo, Kling, Sora, Seedance, Gemini Omni, and every Wan above 2.2 are API-only. We checked each license and repository directly on 2026-08-28, and the checks are reproducible in a browser.
TL;DR
- LTX-2.5 (Lightricks): the pick. Video plus synced audio, text, image, and video inputs, about 1.6 million downloads in the past 30 days, LTX-2.x Community License (no exclusion for the US, UK, EU or South Korea), free commercially under $10M revenue
- Wan 2.2 (Alibaba): Apache 2.0, the permissive pick, 720p at 24fps, 5-second clips, 24GB VRAM
- MiniMax H3 (MiniMax): video with stereo audio, 768p by default, the most downloaded here (about 3.57 million in the past 30 days), but its licence does not cover the US, the UK, the EU or South Korea
- Mochi 1 (Genmo) and HunyuanVideo (Tencent): real open weights, but 2024 to 2025 generation quality
- Not open despite what you will read: Wan 2.5, Wan 3.0, Veo 3.1, Kling v3, Sora 2, Seedance 2.5, Gemini Omni
- If you do not own a 24GB-class GPU, hosted is cheaper than the hardware until you are running thousands of clips a month
The list
1. LTX-2.5
License: LTX-2.x Community License Agreement. Commercial and production use at no cost under $10M revenue; above that, a paid commercial agreement is required.
The catch: the license is not OSI-open, it is revenue-gated. For an indie team or a studio under the threshold that distinction costs nothing. For a large company it is a procurement conversation, not a download.
LTX-2.5 landed on Hugging Face on 2026-07-23 and was still being updated on 2026-08-27, with 912,729 downloads when we first checked on 2026-08-28. The repository tags tell the real story about scope: text-to-video, image-to-video, video-to-video, audio-to-video, text-to-audio, video-to-audio, and the combined text-to-audio-video and image-to-audio-video paths. It ships a separate audio VAE and generates picture and sound together, which is the capability that separated closed 2026 models from open 2025 ones until this release.
The reference configuration in the model card runs 960x544 at 24fps for 121 frames, about 5 seconds. Frame counts must satisfy num_frames % 8 == 1 and both dimensions must be divisible by 32. The card documents fp8 casting and CPU offload for low-VRAM setups rather than publishing a hard minimum, which in practice means it runs on consumer hardware if you are willing to trade speed.
Best for: anyone who wants current-generation output, including audio, without a per-clip meter.
2. Wan 2.2
License: Apache 2.0, the most permissive license on this list.
The catch: it is a 2025 model, and the Wan line went closed above it. Wan 2.5 and Wan 3.0 have no published weights, so this is where the open branch ends.
Wan 2.2 remains the cleanest license story in open video. The Wan2.2-TI2V-5B card states 720p output at 1280x704 or 704x1280, 24fps, 5-second clips, a 24GB minimum VRAM requirement (an RTX 4090 class card), and generation in under 9 minutes on a single consumer GPU. Larger A14B text-to-video and image-to-video variants sit above it, and the Wan 2.2 Animate line for character animation was still receiving uploads on 2026-08-06.
Output quality is honestly a generation behind. Against Kling v3 or Veo 3.1 on the same prompt, the gap shows in lighting subtlety and in how fine detail survives motion. For b-roll, background plates, and drafting, that gap does not matter.
Best for: projects where the license has to be boring, and for fine-tuning, since the community ecosystem around Wan is the largest in open video.
3. MiniMax H3
License: MiniMax H3 Community License Agreement, dated 2026-08-02. It applies worldwide except the European Union, the United Kingdom, the Republic of Korea and the United States. Commercial products with more than $20M in yearly revenue need MiniMax's written authorization, and any commercial product built on it must prominently display "MiniMax H3".
The catch: if you are in the US, the UK, the EU or South Korea, you cannot use H3 under this licence; there, MiniMax takes applications for a separate licence. The open release is also only part of the system MiniMax offers through its API.
MiniMax published the weights on Hugging Face on 2026-07-28, with no access gate, and they had about 3.57 million downloads in the past 30 days when we checked on 2026-10-02. Per the model card, the open part is H3-Base in two checkpoints: text or first and last frames to video, and reference to video from images, video clips and audio. It outputs video with 32 kHz stereo audio, 24 FPS, with the shorter side at 768 pixels by default, and the card lists clips of 4 to 15 seconds. Two pieces are not in the open release: H3-Context-IR, the prompt pre-processing the card strongly recommends, and H3-Regenerate-2K, the 2K step; both run through MiniMax's API. The card's SGLang serving example uses 4 GPUs.
Best for: teams outside the four excluded territories with multi-GPU hardware who want open weights that take image, video and audio references.
4. Mochi 1 (Genmo)
License: Apache 2.0.
The catch: preview-stage weights from 2024, and no audio.
Mochi 1 is the other genuinely Apache-licensed video model with real adoption. It is worth knowing about mostly as a fallback when a license review rules out revenue-gated terms and Wan does not fit the shot.
5. HunyuanVideo (Tencent)
License: custom Tencent community license, not Apache.
The catch: the license carries regional and use restrictions, and the download numbers on the base repository are small compared to LTX and Wan, which means a thinner ecosystem of tooling and LoRAs.
What is not open, no matter what the listicles say
| Model | Status | Where it runs |
|---|---|---|
| Wan 2.5, Wan 3.0 | No published weights | Alibaba Cloud, fal, Replicate, hosted platforms |
| Veo 3.1 (all tiers) | Closed | Gemini API, Flow, hosted platforms |
| Kling v3 | Closed | Kling API, hosted platforms |
| Sora 2 | Closed, and retired on 2026-09-24 | No longer available |
| Seedance 2.5 | Closed | Replicate, hosted platforms |
| Gemini Omni Flash | Closed | Gemini API, hosted platforms |
If a page tells you one of these is open source, it is repeating a claim nobody checked. The check takes one request: search the vendor's Hugging Face organization and look for the version number.
The hardware math nobody runs before buying
Self-hosting is free per clip and expensive everywhere else. A 24GB card, an inference stack, CUDA version drift, and the hours you spend on all three. The break-even is higher than enthusiasm suggests.
Concretely: a 5-second 720p clip from the model you would self-host costs 50 credits on the 8frame canvas, and the $19 Starter plan carries 1,000 credits a month. That is 20 clips of exactly the same model, plus the closed models on the same canvas for the shots where open weights are not good enough. Price the GPU, the electricity and your setup time against that before buying hardware.
Self-hosting wins when you are running thousands of clips a month, when you need to fine-tune on your own material, or when the footage cannot leave your infrastructure. Those are good reasons. "It is free" is not one of them until the volume is real.
FAQ
Can I use open-weights output commercially? Depends on the license, and they differ. Apache 2.0 (Wan 2.2, Mochi 1) is permissive. The LTX-2.x Community License allows commercial use at no cost under $10M revenue and requires a paid agreement above it. The MiniMax H3 Community License does not cover the US, the UK, the EU or South Korea at all; elsewhere, commercial products must display "MiniMax H3" and need written authorization above $20M in yearly revenue. Tencent's license carries its own restrictions. Read the card for the exact repository you download.
Which open model is closest to closed-model quality? We have not benchmarked it. Two open releases on this list generate synchronized audio, the feature that defined the 2026 closed models: LTX-2.5 and MiniMax H3 (H3 only outside the US, the UK, the EU and South Korea).
Is there an open model with no license restrictions at all? Wan 2.2 under Apache 2.0 is the closest thing. That is the reason to pick it over a newer model with better output.
Do I still need a hosted platform if I self-host? Often both: open weights for volume and iteration, hosted closed models for the shots that ship. Running the same prompt through both is the fastest way to find out where your own line is. See best AI video generator 2026 for how the closed field compares.
We host Wan 2.2 at 50 credits a clip (on 8frame it runs from a start image or an input video, not from text alone; text-to-video there is Wan 3.0, 24 credits for 5 seconds at 480p) and have not added LTX to the canvas yet, so if you want LTX-2.5 today the route is your own GPU or a raw inference provider. For everything else, run the field side by side on 8frame from $19/month. Related reading: is Wan 3.0 open source and open weights vs closed models.
Sources
- LTX-2.5: Lightricks model card and repository on Hugging Face (huggingface.co/Lightricks/LTX-2.5), licence, tags, reference configuration and download count, checked 2026-08-28; licence name and the $10M threshold re-checked on the card and in the LTX-2.x Community License Agreement, checked 2026-10-01
- LTX-2.5 downloads in the past 30 days (about 1.6 million): Hugging Face API, Lightricks/LTX-2.5, checked 2026-10-02; the LTX-2.x Community License Agreement has no territory exclusion beyond sanctions compliance, checked 2026-10-02
- MiniMax H3: model card (open checkpoints, 768p default, 32 kHz stereo audio, 24 FPS, 4 to 15 s, H3-Context-IR and H3-Regenerate-2K not in the open release, 4-GPU serving example), licence (dated 2026-08-02; excluded territories, $20M authorization, "MiniMax H3" display) and Hugging Face API (created 2026-07-28, not gated, about 3.57 million downloads in the past 30 days), checked 2026-10-02
- Mochi 1: genmo/mochi-1-preview on Hugging Face, Apache 2.0, published 2024-10-22, checked 2026-10-01
- HunyuanVideo: tencent/HunyuanVideo on Hugging Face and its license (Tencent Hunyuan Community License; the territory excludes the European Union, the United Kingdom and South Korea), checked 2026-10-01
- Wan open weights: huggingface.co/api/models?author=Wan-AI and the Wan2.2-TI2V-5B card, checked 2026-08-28
- Sora retirement date, 2026-09-24: OpenAI's deprecations page
- 8frame price for Wan 2.2: 8frame tool registry
- Wan 2.2 input requirement on 8frame and the Wan 3.0 price: 8frame video generation code and tool registry, read 2026-10-01
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.