Best open-source video models (November 2026)

The open-source and open-weight video models at a glance

These are the video generation models whose weights you can download today from the developer's own repository. Every figure is the developer's own: their model card, repo or licence. Where a developer doesn't publish something, the table says so rather than filling the gap with a community number.

Best open-source text to video model

Wan 2.2 T2V-A14B is the safe default. It's a mixture-of-experts model: a high-noise expert lays out the early denoising steps and a low-noise expert handles detail, so 14B of its 27B parameters run per step. The official config accepts four sizes (1280 × 720, 832 × 480 and their portrait versions) and defaults to 81 frames at 16 fps, about five seconds. The Wan team ships the prompt-rewriting instructions it uses, which is the closest thing to an official prompt guide.

HunyuanVideo 1.5, open-weight under Tencent's community licence, does text to video at 480p or 720p in 8.3B parameters, with separate super-resolution models that take output to 1080p. Its default clip is 121 frames at 24 fps. Tencent publishes a prompt handbook and turns prompt rewriting on by default; the README warns that turning it off may lower quality.

LTX-2.5 (open-weight) is the pick if you want sound in the same pass or more than 720p. Its default production pipeline renders 1024 × 1536 at 24 fps, and the README gives 3840 × 2176 as the 4K setting. It also generates several connected shots in one pass, which Lightricks calls native multi-shot. LTX-2.3 runs in the same pipelines; Lightricks' hosted API serves it at up to 4K and 20 seconds, while the open repo sets no hard cap.

LongCat-Video (MIT, open-source) covers text to video, image to video and video continuation in one 13.6B model. The continuation training is the point: the official long-video demo chains 11 continuation segments onto a first clip, which the script describes as a one-minute video.

Best open-source image to video model

Wan 2.2 I2V-A14B is the open image-to-video model with the broadest tool support: ComfyUI and Diffusers, plus every community trainer we checked. The aspect ratio follows your start frame, and --size sets the pixel area (480p or 720p). The official config runs 40 steps at guidance 3.5.

HunyuanVideo 1.5 (open-weight) has the most image-to-video variants: 480p and 720p checkpoints, CFG-distilled versions, a 720p sparse-attention version for H-series GPUs, and a 480p step-distilled model that Tencent recommends running in 8 or 12 steps.

MiniMax H3 (open-weight) takes zero, one or two images (first and last frame) in its FL2VA checkpoint, and up to 9 images, 3 videos and 3 audio clips as references in its Ref2VA checkpoint. Either way the output comes with audio.

Wan 2.2 TI2V-5B does both text and image to video in one 5B model, at 1280 × 704 and 24 fps. It's the Wan model to try first on a single consumer GPU.

For character animation rather than a free-form clip, Alibaba released Wan2.2-Animate-2-14B in August 2026 under Apache 2.0. It takes a reference image, a driving video and a prompt, and its distilled checkpoint runs in 10 steps without classifier-free guidance.

Open-weight video models with audio

None of the open-source models generate sound. Three open-weight models generate synchronised audio and video from one model:

• LTX-2.3 and LTX-2.5 (Lightricks). Both cards list text, image and audio inputs, including audio-to-video and video-to-audio. The LTX-2.3 card notes that audio generated without speech "may be of lower quality".
• MiniMax H3. 32 kHz stereo audio at 24 fps video, with dialogue in 11 languages that MiniMax says it supports stably. Clips run 4 to 15 seconds.

The others don't generate sound, but there are audio-driven siblings. Wan2.2-S2V-14B turns a reference image plus an audio track into video, and its length follows the audio. Meituan's LongCat-Video-Avatar does the same for talking characters. Tencent's HunyuanVideo-Foley goes the other way: it takes a finished video and a text description and generates 48 kHz sound effects for it.

Best open-weight video model for low VRAM

Going by what developers publish, HunyuanVideo 1.5 (open-weight) has the lowest official floor: 14 GB of GPU memory, measured with model offloading on, on Linux with an NVIDIA GPU. Wan 2.2 TI2V-5B, the lowest open-source option, is next, at 24 GB (the README names an RTX 4090) with offloading, dtype conversion and the T5 encoder on CPU. The Wan 2.2 A14B models need 80 GB for the official single-GPU 720p command.

Lightricks, Meituan and MiniMax don't publish a VRAM figure. LTX's README suggests --quantization fp8-cast --offload cpu for memory-constrained GPUs, and its full download is about 66 GiB. MiniMax's only hardware example runs H3 across four GPUs.

Our hardware requirements page collects the published figures for every model in the catalogue.

Commercial-friendly licences

This is where the models differ most, and where open-source and open-weight part ways.

If you're a US or EU business, that narrows the field fast: MiniMax H3 is unlicensed in your territory and HunyuanVideo 1.5 is unlicensed in the EU and UK. Wan 2.2 and LongCat-Video, the open-source pair, carry no such conditions. LTX is free until your company reaches $10M in annual revenue. See commercial use of open-source and open-weight models for the image models too.

LoRA training and the ComfyUI ecosystem

Wan 2.2's LoRA ecosystem comes from the community, not the Wan team: the official repo has no training code and points to DiffSynth-Studio, a community project, for LoRA and full training, and to LightX2V for step-distilled checkpoints that cut the step count. ComfyUI and Diffusers support all three Wan 2.2 base models.

Two developers ship their own training code. Lightricks says the LTX dev transformer is fully trainable with its ltx-trainer package (LoRA, full fine-tuning and IC-LoRA), and the LTX-2.5 card says most LoRAs trained on LTX-2.3 run on 2.5 unchanged. Tencent released train.py for HunyuanVideo 1.5 with a LoRA option and asks that you fine-tune with its Muon optimizer.

Among community trainers, musubi-tuner supports Wan 2.1 and 2.2, HunyuanVideo 1.5 and MiniMax H3 (experimental, including joint audio). ai-toolkit lists Wan 2.2, LTX-2.3, LTX-2.5 and MiniMax H3. H3 needs care: its released checkpoints are CFG-distilled, and musubi-tuner's docs warn that a LoRA trained on the plain flow-matching target degrades it, so it offers adapted loss methods instead.

LongCat-Video's README has no training code.

Is Wan 2.7 or Wan 3 open source?

No. As of November 11, 2026, there are no Wan 2.5, 2.6, 2.7 or 3.0 weights in the Wan-AI organisation on Hugging Face or the Wan-Video organisation on GitHub. The newest downloadable Wan releases are still named 2.2: Wan2.2-Animate-2-14B and its distilled variant, published in August 2026.

Both exist as API-only models on Alibaba Cloud. Wan 2.7 launched in 2026 with 2 to 15 second clips and video editing. Wan 3.0 became generally available in August 2026, with clips up to 30 seconds, up to 1080p, native audio and reference inputs of up to 10 images, 5 videos and 5 audio clips. Whether either will get open weights, we don't know, and we haven't found a statement from Alibaba either way. Pages offering "Wan 2.7 open source download" links are not from the Wan team; check that a checkpoint sits under Wan-AI before downloading it.

About leaderboards

We don't publish a ranking, because there isn't a neutral one for open video models. Public preference arenas change week to week and vary in which open checkpoints they include. Developer benchmarks are run on each developer's own prompts. Meituan, for example, reports LongCat-Video slightly ahead of Wan 2.2 T2V-A14B on overall quality in its internal mean-opinion-score test (3.38 against 3.35) and slightly behind it for image to video (3.17 against 3.26). Treat those as the developer's numbers, not an independent result.

In practice, the constraints decide faster than scores do: your territory and revenue against the licence, whether you need audio, your GPU, and whether you'll train LoRAs.

Running them without your own GPUs

Video generation on Windpaint is coming soon. Wan 2.2 image-to-video, Wan 2.2 text-to-video, TI2V-5B, LongCat-Video and HunyuanVideo 1.5 are in the catalogue as coming soon. Wan 2.2 image-to-video's planned price is 4 credits ($0.20) per 5 second 480p clip; see pricing. Image generation is live today.