5 Best Open-Weights AI Video Models to Run Yourself, Ranked: LTX-2.5 Tops Wan 2.2

Lightricks' LTX-2.5 is the best open-weights AI video model creators can run themselves right now, ahead of Alibaba's Wan 2.2. That is the top of SLOP TV's ranking as of September 19, and it follows a busy August: MiniMax published weights for its H3 model on August 3, and LTX-2.5 arrived on August 11 with day-one ComfyUI support.

The field is now pulling in two directions. One camp is getting bigger: H3 has 33 billion parameters, and Sand.ai's MAGI-2 Preview lists 114 billion, about 307GB of checkpoints and eight Hopper-class NVIDIA GPUs in its setup notes. The other camp is staying small enough for one desktop card, with Tencent's HunyuanVideo 1.5 asking for 14GB of GPU memory and Kandinsky Lab's Video Lite targeting 12GB.

"Open weights" means the checkpoints can be downloaded and run locally. It does not mean open source, and it does not mean free for any use, so the licence gets its own line in every entry. The ranking criteria, in order: whether a creator can realistically run the model on the hardware its maker documents; what the licence allows, and where; what comes out (resolution, clip length, audio); and how well the model is supported in ComfyUI and Diffusers. SLOP TV has not run side-by-side tests for this list, so the ranking rests on the makers' own model cards, repos and licence texts. MAGI-2 Preview misses the cut on hardware alone. In all, five open-weights models made it.

5. MiniMax H3 (MiniMax)

MiniMax H3 is the largest open-weights model on this list and, on paper, the most capable. The model card describes a 33-billion-parameter dense, single-stream transformer that generates 4 to 15 second clips at 24fps with 32kHz stereo audio. It ships as two checkpoints: FL2VA for text and first/last-frame prompts, and Ref2VA for reference-driven work, which accepts up to nine images, three video clips and three audio clips. Output defaults to 768p. The H3-Regenerate-2K module that lifts clips to 2K is "not yet open-sourced", and the H3-Context-IR prompt layer stays behind MiniMax's API.

Now, one caveat, and it is a large one. The MiniMax H3 Community License Agreement, dated August 2, grants rights worldwide excluding the United States, the European Union, the United Kingdom and South Korea. It also requires separate written authorization above $20 million in yearly revenue, and commercial products must display "MiniMax H3" prominently. Hardware is the second hurdle: MiniMax's SGLang example uses four GPUs, although the card notes that roughly 13 billion of H3's parameters do not need to be loaded for inference. A separate research project, VDN-H3, published on September 6, says its H3-based variant peaks at 22GB on a 24GB card for a 345-frame clip, under the same licence. H3 suits reference-driven clips with sound, for creators outside those four territories. It has the most features here and the fewest people who can use them.

4. Kandinsky 5.0 Video Lite (Kandinsky Lab)

Kandinsky 5.0 is one of only two open-weights families on this list with a standard permissive licence: the GitHub repo and the Hugging Face cards carry the MIT licence, with no territory clause and no revenue ceiling. The family includes the 2-billion-parameter Video Lite and the 19-billion-parameter Video Pro, both with text-to-video and image-to-video variants, and prompts work in English and Russian. Video Lite's model card dates it to September 29, 2025; Video Pro was opened on November 20, 2025. Size is the defining strength. With offloading and quantization, the repo says generation should work on GPUs with 12GB of memory, a long way down from H3's four-GPU example.

Kandinsky 5.0 Video Lite fits drafts and longer takes on a modest card. It comes in 10-second variants at 768x512, double the clip length of HunyuanVideo 1.5 one rank up. The catch is picture size and sound: Video Lite is standard definition, Video Pro reaches 1280x720 only on 5-second clips, and the documentation does not mention audio. Kandinsky Lab's own H100 timings show the gap between the two tiers: 35 seconds for a distilled 5-second Lite clip and 61 seconds for a distilled 10-second one, against 1,241 seconds for a 5-second HD Pro clip. The repo also reports Video Pro as the top open-source text-to-video model on LMArena on December 12, 2025, but arena rankings move and the news log has no 2026 entries. Kandinsky 5.0 has one of the two easiest licences here and the lowest resolution.

3. HunyuanVideo 1.5 (Tencent)

Tencent's HunyuanVideo 1.5 is the open-weights model built most squarely for a single consumer GPU. The model card lists 8.3 billion parameters, text-to-video and image-to-video at 480p and 720p, and a few-step super-resolution network that takes output to 1080p, which Kandinsky 5.0 cannot match. Clips run 121 frames at 24fps, or about 5 seconds. The stated minimum is 14GB of GPU memory with model offloading enabled. Weights and inference code were released on November 20, 2025, and the most recent entry in the repo's news log, from December 23, 2025, added FP8 inference.

HunyuanVideo 1.5 is strongest at image-to-video on a desktop card. A step-distilled 480p image-to-video model released on December 5, 2025 generates in 8 or 12 steps, and Tencent puts end-to-end generation on an RTX 4090 within 75 seconds. Training code and a LoRA tuning script came out the same day. The catch is the Tencent Hunyuan Community License. Its territory excludes the European Union, the United Kingdom and South Korea (US creators are covered, unlike with H3), it requires a separate licence above 100 million monthly active users, and it bars using outputs to improve competing AI models. HunyuanVideo 1.5 asks for 10GB less memory than the smallest Wan 2.2 model, but it comes with more legal reading.

2. Wan 2.2 (Alibaba)

Wan 2.2 is the oldest open-weights model on this list and still the safest default, for one reason: Apache 2.0. The Wan 2.2 repo went up on July 28, 2025 with ComfyUI and Diffusers integration, and its licence has no territory clause, user ceiling or revenue cap. The lineup covers a text-to-video and an image-to-video model, each a mixture-of-experts design with 27 billion total and 14 billion active parameters at 480P and 720P, plus the dense TI2V-5B at 720P and 24fps. S2V-14B and Animate-14B variants followed, and the family is still growing: the official Wan-AI page on Hugging Face shows Wan-Dancer-14B uploaded in July and Wan2.2-Animate-2-14B in August of this year.

Wan 2.2 is the pick for commercial work that needs a licence nobody has to argue about. TI2V-5B runs on a 24GB card such as an RTX 4090 and, per the repo, produces a 5-second 720P clip in under 9 minutes. The catch is the hardware step above that: the 14B-class models list 80GB of VRAM for single-GPU inference. Clips stay at 5 seconds, and the repo lists nothing like LTX-2.5's joint audio generation for the text-to-video and image-to-video models. Creators waiting for a newer open Wan should note that, as of September 19, the official Wan-Video organisation on GitHub has no repo for any version after 2.2. Wan 2.2 is a clear step up from HunyuanVideo 1.5 on licence, and a step behind LTX-2.5 on what comes out.

1. LTX-2.5 (Lightricks)

LTX-2.5 tops the open-weights field because it is the only model here that pairs a finished-shot feature list with a desktop-sized memory floor. The model card lists a 22-billion-parameter transformer with an audio VAE and vocoder, in full and distilled bf16 checkpoints plus int8 and NVFP4 quantizations. According to ComfyUI's launch post from August 12, native 4K, synchronized audio and frame rates up to 50fps carry over from LTX-2.3. New in LTX-2.5 are a diffusion video decoder, a custom text encoder and native multi-shot generation, where one run produces several connected shots that hold character, environment, lighting and voice across the cuts. The release landed on August 11, according to DataNorth.

LTX-2.5 suits creators who want picture and sound out of one pipeline. Lightricks' card gives no VRAM figure, only fp8 casting and CPU offload as low-memory options. Runpod's deployment guide puts the floor at 16GB for the quantized variants and 48GB or more for full BF16. The headline speed figure, a 10-second 720p clip in 6.8 seconds as reported by DataNorth, was measured on two NVIDIA GB200 GPUs, which is data-centre hardware. The catch is the licence. The LTX-2.x Community License allows commercial use at no cost below $10 million in annual revenue, requires a paid agreement above that, and says transferring fine-tunes may need a paid licence. The card does not state a maximum clip length for the open weights. Against Wan 2.2, LTX-2.5 adds audio, 4K and multi-shot, and gives up Apache 2.0. For most creators under the revenue line, that trade is worth making.

LTX-2.5's weights are available now from the gated Lightricks/LTX-2.5 repository on Hugging Face after accepting the licence, with official workflow templates in ComfyUI 0.32.0.