Open-Weight AI Video Models Released in September 2026

Key takeaways

• No major lab released a new flagship open-weight video model in September 2026. The two big bases, MiniMax H3 and LTX-2.5, both shipped in August, and most of September's releases build on them.

• The main new general-purpose model is DreamX-Creator 1.0, a 7B joint audio-video generator released under Apache 2.0 on September 3.

• Speed was the month's main theme. At least five releases cut MiniMax H3 down to 3 to 8 sampling steps through distillation, sparse attention or linear attention, and one brings it to a 24GB graphics card.

• World models, which generate video you can steer with a camera or keyboard as it plays, arrived from Robbyant, Tencent ARC and others.

• Read the license before you build on anything here. MiniMax H3's license excludes the US, EU, UK and South Korea without separate permission, and every H3-based release inherits that restriction.

What shipped in open-weight video in September 2026

September was a derivatives month. The two open video bases most of the month builds on landed in August: MiniMax open-sourced H3, a 33B model that generates video and stereo audio together, on August 3, and Lightricks published the weights for LTX-2.5, a 22B joint audio-video model, on August 11. Wan 3.0, the next version of Alibaba's Wan line, has stayed hosted-only.

So the month's releases fall into four groups: a handful of new models, world models, a long list of speed-ups and add-ons for MiniMax H3, and a wave of editing adapters for LTX-2.5.

How we checked: a model counts here if its weights were downloadable from the developer's own Hugging Face or GitHub page and the weight files were first uploaded between September 1 and September 30, 2026. We dated each release from the repository's commit history, not from announcements, and checked architecture, license and hardware claims against the model cards and papers on October 2, 2026. Benchmark and speed figures below are self-reported by the developers unless we say otherwise. We did not run the models ourselves.

New models

DreamX-Creator 1.0 (AMap ML): Released September 3, this is the only new general-purpose model of the month. It's a 7B model that turns a first frame and a prompt into video with synchronized audio, using separate video and audio diffusion transformer streams that are coupled through cross-modal attention in the second half of the network. A separate 5B refiner upscales output to 2K in a single step and also works on videos made by other models. It generates 5 seconds at 24 frames per second by default, the weights total about 54GB, and it can run on one GPU with CPU offloading or split across several. It's licensed under Apache 2.0, which allows commercial use without revenue caps. A faster distilled version is on the team's roadmap.

LingBot-Video MoE-DMD 30B-A3B (Robbyant): A new checkpoint in Robbyant's LingBot-Video family, released September 17. It's a mixture-of-experts diffusion transformer with 30B total parameters but only about 3B active per token, which keeps compute closer to a small model's. This version is distilled to 8 sampling steps and handles text-to-video and text-and-image-to-video at 832x480 and 24 frames per second. The family is trained on large amounts of web video plus more than 70,000 hours of embodied, robot-style footage, so it's aimed at physical tasks more than cinematic clips. On the team's test GPU, memory use peaked at about 82GB. License: Apache 2.0.

TBDub (Alibaba TaoLive): Released September 5, TBDub re-animates a speaker's lips in an existing video to match new audio. A 2-step student model is about 14x faster than its 30-step teacher, and Alibaba reports 7.13 frames per second at 512x512 on a single H20. With CPU offloading, the team measured peaks of about 15GB (student) to 19.5GB (teacher) on a workstation card, though it hasn't validated a 24GB consumer GPU. License: Apache 2.0.

World models

World models generate video that responds to input as it plays, such as camera moves or keyboard controls, instead of rendering a fixed clip. September brought several.

LingBot-World 2.0 (Robbyant): The action-controlled world model, built on Wan 2.2, first shipped in July. On September 7 Robbyant added the remaining checkpoints, a 14B bidirectional model and a 1.3B real-time model. Robbyant claims an unbounded interaction length and says its distilled real-time variant streams 720p at 60 frames per second. License: CC BY-NC-SA 4.0, which rules out commercial use.

WorldCrafter (Tencent ARC Lab and Peking University): Built around a 3D-aware memory, this camera-controlled world model keeps a scene consistent when the camera returns to it. Base and few-step Fast versions were released September 21, and Tencent ARC claims minute-long exploration. License: academic use only.

InSpatio-World 1.5 (InSpatio): A 1.3B video-to-video model, released September 24, that turns a single image, four images or a video into an explorable 4D scene, built on Wan 2.1 and billed for real-time use. License: Apache 2.0 for the code, per its GitHub repo.

Built on MiniMax H3

Most of September's releases are add-ons for MiniMax H3. Many are LoRAs (Low-Rank Adaptations, small add-on weights that modify a base model) rather than full models, so you still need H3 itself: about 66GB for the transformer alone, and about 144GB with its text encoder and VAEs.

Faster generation

• VDN-H3 (OpenVDN, September 2): adds a linear "Video DeltaNet" attention branch alongside H3's local softmax attention and runs in 8 steps. OpenVDN reports a 14.5x speedup over 50-step H3, and it runs on a 24GB card with block offloading, peaking at 22GB.

• FastH3 8-Step V2 (Hao AI Lab, September 4): distills H3's text-to-audio-video mode to 8 steps and adds 80% sparse attention. Its model card says quality falls below base H3 on hard motion and some audio.

• TaoMate-H3 (Alibaba TaoLive, September 7): a 3-step LoRA and runtime for low-latency streaming at 480p to 1080p, validated only on high-end H20 servers.

• HyperFlow (Video Rebirth, September 17): covers all three of H3's modes in 8 steps, with a reported speedup of about 3x.

• LongLive-Plug (NVIDIA, September 29): few-step and guidance-distillation LoRAs for H3, with matching versions for Wan 2.1 and Wan 2.2, built to carry over to fine-tunes of each base without retraining.

Control and editing

• Viggle-Animate (Viggle AI, September 4): a 33B fine-tune of H3 that replaces a character across a whole video from a single repainted frame, with no pose or mask needed. Viggle reports 124 frames in 26 seconds on one B200, 6.1x faster than Wan 2.2 Animate.

• Viggle Meridian (Viggle AI, September 15): re-shoots a video or a single image from new camera angles, for orbits and bullet-time effects, at a 768-pixel-class resolution (1344x768 for widescreen). The two-LoRA version followed on September 18.

• MiniMax-H3-Fun-ControlNet-Union-2.0 (Alibaba PAI, September 22): adds eight control types, including pose, depth, canny, scribble and layout, plus inpainting, for clips up to 15 seconds.

• H3-World (Tencent and university partners, September 1): a 65.6M-parameter LoRA that turns H3 into a keyboard-controlled game-style world model.

• XGEN-JING (XGEN Labs, September 17): a first-person interactive model with camera control, object interaction and dialogue, generating video and audio together.

LTX-2.5 editing adapters

Lightricks spent September releasing adapters for LTX-2.5, mostly IC-LoRAs that apply a specific edit to an existing video. They include day-to-night, colorization, deblurring, decompression, clean plates, restoration, detail refinement, layout-to-render and SDR-to-HDR conversion, plus standard LoRAs for slow motion and cinemagraphs. An alpha matte adapter was uploaded in late September and announced on October 2. On September 29 Lightricks added a Native Resolution mode that tiles edits for 4K and 8K footage on a single GPU, and version 1.4.0 of its LTX-2 software added chunked long-video generation. All of them use the LTX-2 community license, which is free for companies under $10 million in annual revenue.

September's releases compared

What stayed closed

Wan 3.0 has not been released as open weights. Alibaba's official Wan organization on Hugging Face has no Wan 3.0 weights, and its newest open video releases are Wan 2.2 spin-offs from July 2026, Wan-Dancer-14B and Wan2.2-Animate-2. A "wan-3-0-video" page that appeared on Hugging Face in September belongs to an unofficial account and contains no weights.

Licenses: what you're allowed to build

September's releases span four kinds of license. Check which one applies before you compare specs.

• Fully permissive: DreamX-Creator, LingBot-Video and TBDub use Apache 2.0 (as does InSpatio-World 1.5's code), which allows commercial use with no revenue cap.

• Non-commercial: LingBot-World 2.0 (CC BY-NC-SA 4.0) and WorldCrafter (academic use only) are for research.

• Revenue-gated: the LTX-2.5 adapters are free for companies under $10 million in annual revenue.

• MiniMax H3 and everything built on it: H3's license excludes the US, EU, UK and South Korea unless MiniMax grants separate permission, requires separate authorization above $20 million in annual revenue, and bans using H3's outputs to train other models. Derivatives inherit those terms, so most of September's headline releases aren't licensed for use in the US, EU, UK or South Korea without MiniMax's approval.

What runs on a consumer GPU

Most September releases were tested on data-center hardware such as B200, H100 and H20 clusters. The exceptions are VDN-H3, which fits a 24GB card with offloading, and TBDub, which peaked at about 15 to 19.5GB with offloading in the team's tests. Lightricks says LTX-2.5's distilled and quantized versions, which the adapters run on, work on GPUs with as little as 12GB. DreamX-Creator supports CPU offloading but its team hasn't published a memory figure.

Frequently asked questions

What is the newest open-weight AI video model?

Among September 2026 releases, DreamX-Creator 1.0 is the main new general-purpose open-weight video model, released September 3 under Apache 2.0. The main open video bases most September releases build on are MiniMax H3 (33B) and LTX-2.5 (22B), both released in August 2026.

Is Wan 3.0 open source?

No. As of October 2, 2026, Wan 3.0 is only available through hosted services. The newest official open-weight Wan video models are July 2026 releases built on the Wan 2.2 line, such as Wan-Dancer-14B and Wan2.2-Animate-2.

What's the difference between open-weight and open-source AI video models?

An open-weight model lets you download and run the trained weights, but the license may still restrict commercial use, territory or revenue. Open source in the strict sense also means a permissive license and usually the training code. Apache 2.0 models such as DreamX-Creator come closest.

Can I use MiniMax H3 models commercially in the US, EU or UK?

Not without permission. H3's license excludes the US, EU, UK and South Korea unless MiniMax grants separate authorization, and every H3-based release inherits those terms.

Can I run September's open video models on my own GPU?

A few of them. VDN-H3 fits a 24GB card with offloading, TBDub peaked at about 15 to 19.5GB in its team's tests, and Lightricks says LTX-2.5's distilled versions run on as little as 12GB. Most of the rest were tested on data-center GPUs.