9 min readUpdated 18 Aug 2026

6 Free AI Video Tools: What They Cost, What They Need, How to Set Them Up

Only 3 of the 6 are free with no strings. Real VRAM floors, the licence fine print (revenue caps, headcount caps, excluded territories), and the install commands for each — verified against the official repos and licence files.

Verified 17 August 2026. Every claim here comes from an official repo, licence file, or vendor page.

Before the list, the part most guides skip: "free" has fine print. Only 3 of these 6 are free with no strings. Read the licence column before you build a business on one.

Tool	Real VRAM floor	Runs on Mac?	Free for commercial use?
Wan 2.2 (TI2V-5B)	24 GB	No	Yes, Apache 2.0
HunyuanVideo 1.5	14 GB with offloading	Unofficial MLX builds	Yes, except EU / UK / South Korea
Open-Sora 2.0	~52 GB	No	Yes, Apache 2.0
LTX-2.5	~24-32 GB (no official figure)	Slowly, CPU fallback	Only under $10M revenue
Remotion	None, CPU only	Yes, macOS 15+	Only for teams of 3 or fewer
Hyperframes	None, CPU only	Yes	Yes, Apache 2.0, no thresholds

If you have a consumer GPU and want the shortest path: HunyuanVideo 1.5 via ComfyUI. If you have no GPU at all: Hyperframes.

1. Wan 2.2 (Alibaba)

Heads up on the version. You'll see "Wan 2.5" and "Wan 2.6" referenced everywhere. Those are closed API products announced in Dec 2025, with no downloadable weights. The newest Wan you can actually run for free is Wan 2.2, from July 2025. Anyone telling you to run Wan 2.6 locally hasn't checked.

Alibaba's open video family. Wan 2.2 ships a 27B Mixture-of-Experts model (14B active) and a smaller 5B hybrid that fits on one consumer card.

Hardware. The A14B model wants 80 GB. Ignore it. The one you want is TI2V-5B, which the README puts at 24 GB (RTX 4090) for 720p. Below that, run GGUF quantised builds through ComfyUI to get onto 8-12 GB, at the cost of visible detail loss and artefacts in fast motion. No official Mac or CPU path.

Easiest route. ComfyUI, not the research repo. Install ComfyUI, then Template Library → Video → Wan2.2. It fetches weights automatically. To try before installing anything, use the Hugging Face Space.

Install (repo route)

bash

git clone https://github.com/Wan-Video/Wan2.2.git
cd Wan2.2
# needs torch >= 2.4.0
pip install -r requirements.txt
pip install "huggingface_hub[cli]"
huggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir ./Wan2.2-TI2V-5B

First clip

bash

python generate.py --task ti2v-5B --size 1280*704 --ckpt_dir ./Wan2.2-TI2V-5B \
  --offload_model True --convert_model_dtype --t5_cpu \
  --prompt "Two cats in boxing gear fight on a spotlighted stage"

Alibaba's own figure: a 5-second 720p/24fps clip in under 9 minutes on a single consumer GPU.

What breaks. The 720p size is 1280*704, not 1280*720. Get it wrong and you get an error or garbage. And flash_attn is the standard install failure, so install it last or skip it.

Licence. Apache 2.0. Commercial use fine, Alibaba claims no rights over your output.

2. HunyuanVideo 1.5 (Tencent)

Two models share this name. The original HunyuanVideo (13B, Dec 2024) needs 45-60 GB. HunyuanVideo 1.5 (8.3B, Nov 2025) is the one to install: smaller, better quality per byte, runs on consumer hardware.

Licence warning, read this one. The Tencent Hunyuan Community Licence "does not apply in the European Union, United Kingdom and South Korea." If you're in those territories, commercial use is unlicensed. Not a grey area, it's stated in the licence.

Hardware. Official minimum is 14 GB with model offloading enabled, Linux, Python 3.10+, NVIDIA CUDA. ComfyUI targets 24 GB for a comfortable native run. GGUF builds get you to 8-12 GB with softer detail and weaker text rendering.

Easiest route. ComfyUI native support. Template Library → search "Hunyuan Video 1.5 T2V". The tutorial below lists the exact files you need.

Install (repo route)

bash

git clone https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5.git
cd HunyuanVideo-1.5
pip install -r requirements.txt

Weights are at tencent/HunyuanVideo-1.5, in subfolders by resolution and mode (480p_t2v, 720p_i2v, and so on). The repo's checkpoints-download.md has the full list.

First clip. Run generate.py with your prompt. Add --enable_step_distill for the 8-12 step distilled model. Tencent's number: on an RTX 4090, the 480p step-distilled image-to-video model produces a video in about 75 seconds.

What breaks. Linux only, officially. Windows users go through ComfyUI or WSL. The optional accelerators (SageAttention, flex-block-attn, sgl-kernel) each compile from source and are where most installs die, so skip them on your first run.

Licence. Tencent Hunyuan Community Licence, not OSI open source. Commercial use allowed worldwide except EU, UK, South Korea. If your product had over 100 million monthly active users at release, you need a separate licence. You cannot use the outputs to train other models.

3. Open-Sora 2.0 (hpcaitech)

Status: dormant, and not for consumer hardware. Shipped March 2025, last real commit March 2025. Still Apache 2.0, still works, but it's been overtaken on quality-per-VRAM by everything else here. Included because it's genuinely, unconditionally free and the full training code is public, which matters if you want to fine-tune rather than just generate.

An 11B open text-to-video model, notable for being trained for roughly $200K.

Hardware. This is the datacenter one. The repo's own numbers on H100 class hardware: 256x256 peaks at 52.5 GB and takes about a minute; 768x768 peaks at 60.3 GB and takes 28 minutes. There's no quantised consumer build and no ComfyUI workflow. Realistically you rent an 80 GB GPU or you skip it.

Try it without installing: the Hugging Face Space.

Install

bash

conda create -n opensora python=3.10 && conda activate opensora
git clone https://github.com/hpcaitech/Open-Sora && cd Open-Sora
pip install -v .
pip install xformers==0.0.27.post2 --index-url https://download.pytorch.org/whl/cu121
pip install flash-attn --no-build-isolation
huggingface-cli download hpcai-tech/Open-Sora-v2 --local-dir ./ckpts

First clip

bash

torchrun --nproc_per_node 1 --standalone scripts/diffusion/inference.py \
  configs/diffusion/inference/t2i2v_256px.py --save-dir samples \
  --prompt "raining, sea" --offload True

What breaks. That pinned xformers version is built for CUDA 12.1. On any other CUDA version you must swap in the matching wheel, or the install silently mismatches and crashes at inference. Also num_frames must equal 4k+1 and stay under 129.

Licence. Apache 2.0, including the training code. No restrictions.

Repo

4. LTX-2.5 (Lightricks)

Version note. The repo is called LTX-2 but the current model is LTX-2.5, released 11 August 2026. LTX-2.3 is legacy and the checkpoints are not interchangeable.

Revenue cap. Free for individuals and companies under $10M annual revenue, measured across all subsidiaries and affiliates together. Above that you buy a Commercial Use Agreement, and Lightricks charges back-fees for unlicensed use.

A 22B model that generates synchronised video and audio in one pass, including dialogue and lip sync, up to 4K and 50fps. The audio-in-one-pass part is what makes it worth the setup.

Hardware. Lightricks publishes no official VRAM floor. I checked the README, the install doc, and the model card. What is documented: the bf16 weights are about 66 GiB to download, and there are --quantization fp8-cast --offload cpu flags for constrained GPUs. ComfyUI's template defaults to an int8 build, which tells you the consumer path. Treat 24 GB with the quantised ComfyUI build as the practical entry point, and note that any specific number you see quoted elsewhere is community-derived, not official.

Easiest route. ComfyUI, native since 0.32.0. Template Library → Video → LTX-2.5. You must accept the licence on the gated Hugging Face repo first or downloads fail.

Install (repo route)

bash

git clone https://github.com/Lightricks/LTX-2.git && cd LTX-2
uv sync --extra natten     # natten is Linux + CUDA only, auto-skipped elsewhere
hf auth login
hf download Lightricks/LTX-2.5 --local-dir models/ltx-2.5

Note it uses uv, not pip.

What breaks. The Hugging Face repo is gated, so a 401 or 403 on download means you haven't accepted the terms or your token lacks the "read gated repos" scope. That's the single most common failure. Also the bundled Gemma 4 12B text encoder is not interchangeable with Google's stock Gemma 4; the loader checks the version string and rejects mismatches.

Licence. LTX-2.x Community Licence, not OSI open source. Free commercially under $10M revenue only. No training competing models, no removing the watermarking or provenance features.

5. Remotion

Read the licence before you build on this. Free for individuals, non-profits, and for-profit companies with up to 3 employees. At 4 people you need a paid Company Licence. Plenty of teams ship on the free licence and are out of compliance the day they make their fourth hire.

Not an AI model at all. It's a React framework: you write components, Remotion renders them frame by frame in headless Chrome and encodes with FFmpeg. No GPU, no weights, fully deterministic. Best tool here for repeatable, data-driven video, the kind where you generate 500 variants from a spreadsheet.

Hardware. CPU only. Node 16+. macOS 15 or later is required, older macOS simply won't work. Linux needs glibc 2.35+; Alpine and NixOS are unsupported.

Install

bash

npx create-video@latest --yes --blank my-video
cd my-video && npm i
npm run dev      # opens Remotion Studio

First render

bash

npx remotion render

Pricing if you're over the threshold. Automators tier is $0.01 per render with a $100/month minimum. Creators is $25/month per seat. Enterprise starts at $500/month.

If you want the full walkthrough of driving Remotion with a coding agent, I wrote one: Website to video with Claude Code + Remotion.

6. Hyperframes (HeyGen)

The only one on this list with no catch. Apache 2.0, no revenue thresholds, no per-render fees, no employee limits. Worth saying plainly because it sits right next to Remotion, which has all three.

An open-source HTML-to-video renderer. You write a composition as an index.html with timing attributes, and it seeks a headless Chrome page frame by frame and encodes with FFmpeg. Same idea as Remotion, but plain HTML instead of React, and built to be driven by coding agents.

Hardware. CPU only. Node.js 22+ and FFmpeg on your PATH. Works on macOS, Linux, Windows.

Install

bash

npx hyperframes init my-video
cd my-video
npx hyperframes preview    # live-reload preview
npx hyperframes render     # outputs MP4

If you use Claude Code, Cursor, or Codex, this is faster than hand-authoring:

bash

npx skills add heygen-com/hyperframes --full-depth

Then just ask: "Using /hyperframes, create a 10-second product intro with a fade-in title, background video, and subtle music."

What breaks. Node 22 is a hard floor, Node 20 fails. FFmpeg is not bundled, so it must already be on your PATH. And keep --full-depth on skills add, otherwise you get a registry snapshot that lags behind main.

One more thing on the repo. Search results surface hyperframes/hyperframes, which is a near-empty fork. The real one is heygen-com/hyperframes.

Which one should you actually use

No GPU: Hyperframes, or Remotion if you're a team of 3 or fewer and prefer React.

Consumer GPU, 12-24 GB: HunyuanVideo 1.5 through ComfyUI. Best quality you can run at home, as long as you're not doing commercial work in the EU, UK, or South Korea.

You want audio generated with the video: LTX-2.5, if you're under $10M revenue.

You want zero licence restrictions ever: Wan 2.2 or Hyperframes. Both Apache 2.0, no thresholds, no territories.

You want to fine-tune on your own data: Open-Sora 2.0, the only one that published full training code under Apache 2.0. You'll need rented datacenter GPUs.

Built something with one of these? Send it to me, I want to see it.

@gracefrank.ai