Open Text-to-Video Models on Hugging Face Add Native Audio in 2026

A guide that aged in seven months

In early March 2026, a widely shared "complete guide" to text-to-video models on Hugging Face named four top picks: Alibaba's Wan2.2-TI2V-5B, Tencent's 13-billion-parameter HunyuanVideo, a GGUF-quantized build of Wan2.2-T2V-A14B, and Alibaba's image-to-video model I2VGen-XL.1 The guide said open-source models were starting to challenge commercial tools such as Runway and Luma. It also said quantization was cutting the hardware needed to run them, taking VRAM requirements from more than 60GB to under 10GB.1 Its comparison table rated Wan2.2-TI2V-5B as the best option for an RTX 4090. It recommended HunyuanVideo for studios with 80GB A100s and the GGUF build for cards with less than 12GB of memory.1 Copies of the same piece appeared on several aggregator sites within minutes of the original.345

By mid-October, the advice is still useful, but the guide misses where the field has moved. The biggest change since March is not sharper video. It is that leading open models now produce sound as well as pictures. The March guide never mentions audio, and by this autumn that has become the main thing separating one open model from another.

What the Hugging Face listings show now

Hugging Face's own model listings show the change most clearly. In the image-text-to-video category, MiniMaxAI's MiniMax-H3, a 33B model, has about 3.77 million downloads and more than 6,000 likes. An Unsloth GGUF version of it has around 1.82 million downloads.7 The text-to-video trending list is filled with H3 add-ons: turbo LoRAs, camera-motion LoRAs, NVFP4 and INT8 quantizations, and ComfyUI conversions.26 Older models such as Genmo's Mochi-1 preview still appear, but they are far outnumbered by the tooling built around H3.6

H3 matters because it is not only a video model. It is described as a 33-billion-parameter "omni-modal" system. It takes in text, images, video and audio, and it produces video with synchronized stereo sound in one pipeline.17 MiniMax launched it as an API on July 31, 2026, and released the base weights on Hugging Face in August as two checkpoints. One generates video from a first and last frame. The other accepts up to nine reference images, three video clips and three audio clips, and produces 4- to 15-second clips at up to 2K with stereo audio.20 Reports also say it supports dialogue in 11 languages.23

The fast spread of quantized versions repeats a pattern the March guide spotted. Then it was GGUF builds of Wan2.2 that let people run large models on small GPUs.1 Now the community is doing the same thing for H3.

Sound is the new dividing line

The sources agree that synchronized audio is now the main way to tell open video models apart. Lightricks released LTX-2's weights, inference package and training tools in January 2026. It now recommends LTX-2.5 over both LTX-2 and LTX-2.3.25 LTX-2.3, released March 5, is described as a 22B model that pairs a 14B video stream with a 5B audio stream. It outputs up to 4K at 50fps for clips of up to 20 seconds.11 LTX-2.5 appeared on Hugging Face on July 23. One reviewer counted more than 912,000 downloads by late August.26 Lightricks says generating audio and picture in a single pass keeps dialogue and sound effects from drifting out of sync. The company can also generate video from an existing audio track and replace dialogue in footage.13

This is where the coverage disagrees. A GPU-rental company's October comparison still calls LTX-2.3 the only open-source model with native audio.11 That is hard to square with H3's release.17 It also conflicts with reports that Alibaba's Wan 2.7, released in April, includes native audio and voice cloning under an Apache 2.0 license.1218 A third account cuts against those Wan claims: as of September, the official Wan-AI organization on Hugging Face still offered nothing newer than version 2.2.20 The same pattern shows up for Wan 3.0. One ranking calls it open-weight and first on the Artificial Analysis text-to-video leaderboard.19 Another lists it with per-second API pricing on Alibaba Cloud.23

Our reading: anyone planning to run models on their own hardware should go by what is actually downloadable. On that basis, Alibaba's latest open release is still Wan 2.2. Wan 2.7 and 3.0 are best treated as hosted products until weights show up on Hugging Face. On that same basis, H3 and LTX-2.5 are the two open models that generate audio.

Voice synthesis moves inside the video model

This matters for AI voice tools too. When a video model can lip-sync dialogue in many languages, the usual workflow of generating silent video, recording narration and syncing it in an editor becomes optional. One local-video guide still recommends that route for silent models: produce the video, then add voices with Bark, MeloTTS or a local voice clone, and sync in DaVinci Resolve, CapCut or Premiere.22 A game-development guide recommends standalone open voice models such as Qwen3-TTS and Chatterbox, and Stable Audio 3.0's Small-SFX model for sound effects.20

We expect both approaches to coexist for a while. Separate voice models give tighter control over a voice and do not require huge GPUs. Built-in audio solves the sync problem, but it comes with larger models and stricter licenses.

Image generation is converging too

The same convergence is reaching image generation. Black Forest Labs, best known for its FLUX image models, announced FLUX 3 on July 23 as a single model for images, video and audio.20 It offers up to 20-second clips with native sound, keyframe control and lip-synced dialogue in more than 14 languages.12 As of September it was only available through an API. An open-weight FLUX 3 Dev was planned for later in 2026.20 If that release happens, a top image-model lab would be putting an open model directly into the same space as H3 and LTX-2.5.

Licensing: the messiest part of the coverage

If audio is the main technical difference, licensing is the one that most affects who can use these models commercially. The sources agree on Wan 2.2: it is Apache 2.0, with no revenue caps or regional limits.2721 One reviewer said that is still the reason to choose it over newer models that produce better output.26 Lightricks' license is free for organizations under $10 million in annual revenue and requires a paid license above that.2521

The other licenses are less clear. HunyuanVideo 1.5 is variously described as excluding the EU, UK and South Korea25, as allowing commercial use below 100 million monthly active users21, and in one March comparison simply as Apache 2.0.24 For H3, one source says the license excludes the US, EU, UK and South Korea.23 Another describes it as a community license for companies under $20 million in revenue.27 One roundup even calls LTX-2 Apache 2.0,29 which other sources contradict.25 The March guide's advice to check each license before using a model commercially1 matters even more now. For many Western companies, the most popular model on Hugging Face may not be legally usable at all.

Hardware got better and worse at once

The most practical point in the March guide still holds: quantization makes large models usable on consumer hardware. HunyuanVideo 1.5, an 8.3B model, is now listed as running on about 14GB of VRAM with offloading.20 One source says it can render in 75 seconds on a single RTX 4090.23 Wan 2.2's 5B model is listed as running from 8GB in GGUF form.21 LTX-2.5 is reported to run from 12GB.23

The audio-capable models are the exception. Running LTX-2.3 at full BF16 precision is listed as needing 80GB,21 and one guide describes H3 as cloud-first in practice.16 Several hosted services now offer open-weight models for cents per second. One sells LTX-2.5 at $0.01 per second and argues that renting is cheaper than owning a GPU.27 So the open-weight model now covers more than local use. The same weights also feed a growing market of resellers offering them as APIs.

What it adds up to

Put together, the coverage points to three conclusions. First, the March list was accurate for silent video, and Wan 2.2 is still the safest choice when a fully permissive license matters most.2126 Second, the cutting edge on Hugging Face has moved to models that handle several media types and produce sound, led by H3 and LTX-2.5.725 Third, claims about which models are "open" vary more from one source to the next than claims about quality. Readers should confirm a model's weights and license on Hugging Face itself before relying on any ranking.