AI Video Models on a Free T4 GPU: What Fits and What Does Not
Free Colab and Kaggle notebooks usually come with a T4, a GPU with about 15 GB of memory. Many open AI video models were built for much bigger cards. This article lines up four of them, Wan (through Wan2GP), CogVideoX, LTX-Video and HunyuanVideo, against that 15 GB limit.
One thing first: we have not benchmarked these ourselves; every figure below is from the project’s own page, linked. We read each page on 10 November 2026. Models change fast, so check the links before you start a long download.
Which open video models fit a free T4?
Short answer: one Wan setup, CogVideoX with memory-saving tricks, and nothing documented for LTX-2 on a T4 itself. The original HunyuanVideo does not fit.
| Model | What the project says about memory | Fits a free T4? | Route documented to work | Source |
|---|---|---|---|---|
| Wan (via Wan2GP-on-Colab) | "free Colab GPU runtimes usually provide only 15 GB of T4 VRAM", which is "too small for most Wan2GP checkpoints" | Yes, but only for one model: Wan 2.2 TextImage2Video 5B FastWan at 480p | Use that model and select 480p; the changelog entry dated 2026-08-07 says "Tested on a free-tier Colab T4." | Wan2GP-on-Colab README, read 10 Nov 2026 |
| CogVideoX-5b | "diffusers BF16: starting from 5GB" and "diffusers INT8(torchao): starting from 4.4GB"; without the offload and tiling options, peak VRAM is "about 3 times higher than the table shows" | Only with quantization or offload | Quantize and call `pipe.enable_model_cpu_offload()` and `pipe.vae.enable_tiling()`; the card says this makes it possible to run on "free-tier T4 Colab" | CogVideoX-5b model card, read 10 Nov 2026 |
| LTX-Video (LTX-2) | A GitHub issue says LTX-2 "currently requires Gemma 3 12B as the text encoder, which needs ~24-27GB VRAM to operate" | Not documented for a T4. Documented on a 16 GB card only with quantization and offload | An FP8 Gemma file plus ComfyUI’s `--novram` flag, reported to work on a 16 GB RTX 5080 with 64 GB of system RAM | ComfyUI-LTXVideo issue 303, replies dated 6-9 Jan 2026, read 10 Nov 2026 |
| HunyuanVideo | "The minimum GPU memory required is 60GB for 720px1280px129f and 45G for 544px960px129f" | No | None given on the model card | HunyuanVideo model card, read 10 Nov 2026 |
"Fits" here means the source says it runs. It does not say the result is good. Keep reading for a case where it ran and the output was empty.
What does Wan2GP say about a free T4?
The Wan2GP-on-Colab README has a warning at the top. It says free Colab GPU runtimes usually provide only 15 GB of T4 VRAM, which is too small for most Wan2GP checkpoints. Its advice is to stick to the `Wan 2.2 TextImage2Video 5B FastWan` model and select 480p to run without memory errors.
That is a narrow route: one model, one resolution. The README also says the free tier works and paid tiers provide more VRAM. The changelog entry dated 2026-08-07 says "Tested on a free-tier Colab T4." We have not repeated that test, so we cannot tell you how long a clip takes or how good it looks.
If you want the notebook basics first, see our Colab GPU limits guide.
What does CogVideoX say about memory?
The CogVideoX-5b card lists memory for single-GPU runs. In the diffusers library it shows BF16 "starting from 5GB" and INT8 through torchao "starting from 4.4GB". It also lists "SAT BF16: 26GB", a different code path that would not fit a 15 GB card.
Those low numbers depend on switches. The card says that without the offload and tiling optimizations, peak VRAM is about 3 times higher than its table shows. The two calls it points to are `pipe.enable_model_cpu_offload()` and `pipe.vae.enable_tiling()`. With quantization on top, the card says this "makes it possible to run the model on free-tier T4 Colab or smaller VRAM GPUs as well!"
One caution from the same card: it says the VRAM figures have not been tested on devices other than NVIDIA A100 and H100. The T4 is not among the tested devices, so treat the 5 GB and 4.4 GB figures as a starting point only.
What happened when someone tried CogVideoX-2b on a free T4?
A Hugging Face discussion titled "Unable to run on free tier Colab" (opened 28 Aug 2024, closed 30 Aug 2024) shows both sides. A user on a free-tier T4 with 8-bit quantization got "OutOfMemoryError: CUDA out of memory. Tried to allocate 35.31 GiB." A maintainer replied with three calls: `enable_model_cpu_offload()`, `vae.enable_slicing()` and `vae.enable_tiling()`.
The user ran the code and then reported that "for some reason the output file was blank." That is the reminder to keep: a model that runs without an error is not the same as a model that gives a usable clip. Source: CogVideoX-2b discussion 10, read 10 Nov 2026. It is about the 2B model and is more than a year old, so it shows the pattern rather than today’s behavior.
What about LTX-Video?
The LTX evidence is about LTX-2 and a 16 GB card, not a T4. A GitHub issue opened on 6 Jan 2026 says LTX-2 needs Gemma 3 12B as its text encoder, which needs about 24-27 GB of VRAM. The post called LTX-2 unusable on consumer GPUs with 16 GB.
Replies the same day describe two workarounds. One is an FP8 version of the Gemma encoder, placed in ComfyUI’s text encoder folder. The other is ComfyUI’s `--novram` flag, which the thread says streams weights from system RAM. One user reported running it on an RTX 5080 with 16 GB of VRAM and 64 GB of RAM. The issue was closed as completed the same day.
Notice what is missing. Nobody in that thread used a T4, and the setups had far more system RAM than a typical free notebook, so we do not claim it fits. Check the repo before you plan around it, because this area moves fast. We did not read LTX’s own system requirements page, which did not load for us, so we are quoting only the issue.
Why does HunyuanVideo not fit?
The model card is blunt: the minimum GPU memory required is 60 GB for 720 by 1280 pixels at 129 frames, and 45 GB for 544 by 960 at 129 frames. It recommends a GPU with 80 GB of memory for better generation quality. Against 15 GB, that is a clear no for the original model.
The model page does list Colab and Kaggle among its notebook options, but the card itself gives no setup for a small GPU. Community-made smaller versions may exist, but we did not verify any, so we make no claim about them.
How should I choose a model for a free notebook?
Work from the table above, not from a headline. A short checklist:
- Start with the one route a project says it tested on a free T4. For Wan2GP that is the 5B FastWan model at 480p.
- If you pick CogVideoX, turn on offload and tiling from the first run, and expect to troubleshoot.
- Check the output, not just the log. A finished run can still give a blank or poor video.
- Write results to Google Drive as you go, since free sessions can end early (our Colab GPU limits guide covers this).
- Expect download size to matter. Cache model files in Drive once rather than pulling them every session, as that guide explains.
Kaggle is the other free option. Its free GPU is also the T4 (T4 x2 per its docs, two cards of about 15 GB each, as our Kaggle guide covers), so the per-card verdicts in the table apply there too; none of the sources above test Kaggle specifically. Kaggle shows a weekly GPU quota on its settings page. Our Kaggle GPU free limit guide covers it, including the terms. Whichever platform you use, read its current rules first, because both change them.
FAQ
Q: Can I run Wan 2.2 on a free Colab T4?
The Wan2GP-on-Colab README says the free tier’s 15 GB T4 is too small for most checkpoints, and points to Wan 2.2 TextImage2Video 5B FastWan at 480p. Its changelog says that setup was tested on a free-tier T4 on 2026-08-07.
Q: Does CogVideoX run on a free T4?
Its model card says quantization makes that possible, with CPU offload and VAE tiling turned on. The card also says memory was not tested on cards other than A100 and H100, and a 2024 thread shows a T4 user ending up with a blank video.
Q: Can LTX-Video run in 15 GB?
We found no source for a T4. The only documented route is for LTX-2 on a 16 GB card, using an FP8 text encoder and ComfyUI’s `--novram` flag with plenty of system RAM.
Q: Is HunyuanVideo possible on a T4?
Not the original model. Its card gives a 45 GB minimum for 544 by 960 at 129 frames.
Q: Did you test these models?
No. We have not benchmarked them. Every figure here comes from the linked project pages, read on 10 November 2026.
Need a finished product video instead?
If you would rather not fight GPU memory, tell us what you want made and we will take it from there: tell us what you want to make.
Email us your brief
Email us your brief and we reply with a quote.