Open-Source Runway Alternative (2026, Free)

Listed models: Hy-Nomas (Tencent Apache-2.0 interactive world/video DiT, text + camera, 126 likes / 431 GitHub stars, added 2026-11-01; not Hy-Memory), LTX-2.5 (pixel-direct 22B audio-video DiT), Pika H3 Flex (post-trained MiniMax H3 adapters; local 480p/768p open weights — 1080p/2K/4K and audio stems are Pika-hosted; MiniMax H3 Community License; 166 likes, added 2026-10-30; compare vs SenseNova 6.8 Code2Video), and SenseNova 6.8 Code2Video (SenseTime Apache-2.0 27B code-driven short video, Qwen3.8-27B fine-tune; not photorealistic footage). Related still-image generation lives on Midjourney-class alternatives. Pixel editors are on Photoshop alternatives. HTML→MP4 for agents: HyperFrames on HeyGen alternatives.

Sourced snapshot (2026-10-13)

Field	LTX-2.5
Hosts	huggingface.co/Lightricks/LTX-2.5 (gated); github.com/Lightricks/LTX-2
HF likes / downloads	7,589 / 1,724,548
GitHub stars / forks	9,727 / 1,544
Transformer	ltx-2.5-22b-dev and ltx-2.5-22b-distilled
Encoder	Gemma 4 12B bundled (gemma4-12b-ltx-v1)
License	other / LTX-2.x Community License Agreement (11 Aug 2026); $10M revenue clause in §2.1
Distilled download	“roughly 66 GiB” (README)
Quality vs Runway	Not independently tested

Why LTX-2.5 is on this hub

People who search “open source Runway alternative” usually want local, downloadable video generation with a readable license. On 2026-10-13 we opened the public Hugging Face API for Lightricks/LTX-2.5 (7,589 likes, created 2026-07-23, gated, license label other) and the public GitHub repository Lightricks/LTX-2 (9,727 stars, 1,544 forks). The README describes a DiT audio-video foundation model and names LTX-2.5 as the recommended checkpoint, with ltx-2.5-22b-distilled and ltx-2.5-22b-dev transformer files plus a bundled Gemma 4 12B encoder.

Runway itself is a commercial video and “world model” studio. We use its name and mark only to identify that search. We did not run LTX against Runway Gen, Act-Two, or any other SKU.

What “local” means here

The official quick start clones GitHub, downloads roughly 66 GiB of distilled components, and runs python -m ltx_pipelines.distilled. A slower DFR path uses the same distilled transformer plus a detailing IC-LoRA. Both are local once the gated Hugging Face terms are accepted. There is also a hosted LTX playground linked from the README (console.ltx.video/playground). Hosted LTX is not the same as the MIT-style “do anything” story some searchers expect: the community license is not Apache-2.0 or MIT.

License, in one paragraph

Hugging Face will not give you a simple SPDX id. The linked LTX-2.x Community License Agreement (11 August 2026) says Entities with annual revenues of at least $10,000,000 need a paid license except for a defined Non-Commercial Purpose, and should contact [email protected]. If you are evaluating for a company, read the file. forkalt does not interpret whether you are a Commercial Entity.

What we left out

MiniMax-H3 (MiniMaxAI/MiniMax-H3) remains without its own identification page. Pika H3 Flex is the post-trained adapter pair we opened (MiniMax H3 Community License; local 480p/768p). That is not a substitute for a MiniMax-H3 base-card page.

We also do not list closed APIs, unofficial Gradio wrappers, or ComfyUI workflows we have not run.

How to choose (without a quality score)

• Need an identification URL and the actual LTX-2.5 files: start at LTX-2.5.
• Need a still-image generator instead of video: Midjourney-class models.
• Need a chat model, not a video DiT: ChatGPT alternatives.
• Need HTML→MP4 for agents, not a generative DiT: HyperFrames on HeyGen alternatives.
• Need code-driven short video (motion graphics / explainers; not photorealistic pixels): SenseNova 6.8 Code2Video.
• Need local 480p/768p pixel/audio-video adapters on MiniMax H3 (4K/stems hosted): Pika H3 Flex. Compare Pika H3 Flex vs SenseNova 6.8 Code2Video.
• Need ungated Apache-2.0 interactive world/video with camera control: Hy-Nomas. Not Hy-Memory.

What “open” does not mean

LTX-2.5 is downloadable after you accept a gated card, and the inference code is on GitHub. It is not MIT or Apache-2.0. The community agreement (11 August 2026) is the document that matters for companies. The distilled quick start is documented at roughly 66 GiB of files plus a Gemma 4 12B encoder that is not interchangeable with Google’s stock Gemma 4 release (gemma4-12b-ltx-v1 check in the README). If a blog post tells you to “just load Gemma 4,” that is not what the official install says.

Runway’s homepage OG image on this hub identifies the commercial search, not a claim that LTX clones that UI. For still images, do not start here: use Midjourney-class alternatives. For speech-to-notes, use Otter.ai alternatives. For chat, use ChatGPT alternatives.

Paper linked from the LTX-2 README: arxiv.org/abs/2601.03233. We have not summarized that paper on this hub. Discord and website badges on the README are community links, not extra license grants.