RTX 3060 12GB AI Video Models: What Runs, What Fits, What Was Measured
Which open AI video models run on an RTX 3060 12GB in ComfyUI: Wan, HunyuanVideo 1.5, LTX, CogVideoX, Mochi and MiniMax H3, with file sizes, official figures and the system RAM they need.
Quick answer
An RTX 3060 12GB can run most of today's open video models in ComfyUI. Very few fit inside its 12 GB. ComfyUI keeps what it can on the card and streams the rest from system RAM, so on this card system RAM decides more than VRAM does.
We measured that on our own RTX 3060 with MiniMax H3: the card peaked at 11,649 of 12,288 MiB, and system memory peaked at 43,587 MiB. A 32 GB machine would not have held it.
| Model | Smallest common ComfyUI file | Fits in 12 GB alone? | Measured on an RTX 3060? |
|---|---|---|---|
| ModelWan 2.1 1.3B | Smallest common ComfyUI file2.84 GB, fp16 | Fits in 12 GB alone?Yes | Measured on an RTX 3060?No published run |
| ModelWan 2.2 5B | Smallest common ComfyUI file10.00 GB, fp16 | Fits in 12 GB alone?Yes, with little left over | Measured on an RTX 3060?No published run |
| ModelHunyuanVideo 1.5, 480p | Smallest common ComfyUI file8.33 GB, fp8 | Fits in 12 GB alone?Yes | Measured on an RTX 3060?No published run |
| ModelWan 2.1 and 2.2 14B | Smallest common ComfyUI file14.29 GB, fp8, per file | Fits in 12 GB alone?No; a Q4_K_M GGUF is 9.65–10.12 GB | Measured on an RTX 3060?No published run |
| ModelMochi preview | Smallest common ComfyUI file10.03 GB, fp8 | Fits in 12 GB alone?Yes, with little left over | Measured on an RTX 3060?No published run |
| ModelCogVideoX 5B | Smallest common ComfyUI file11.14 GB, transformer | Fits in 12 GB alone?Barely | Measured on an RTX 3060?No published run |
| ModelLTX-2.5 | Smallest common ComfyUI file21.50 GB, int8 | Fits in 12 GB alone?No | Measured on an RTX 3060?No; one run on a 12 GB RTX 4070 |
| ModelMiniMax H3 | Smallest common ComfyUI fileSeveral parts; see below | Fits in 12 GB alone?No | Measured on an RTX 3060?Yes: ours, and one other owner's |
"Fits" here means the file is smaller than the card. It is not a VRAM measurement: the model also needs working memory for your resolution and frame count, and the text encoder and VAE need room in turn.
We read every official statement, file size and third-party page on 2026-10-09. File sizes are from the Hugging Face API, in decimal gigabytes. The only RTX 3060 video measurements on this page are for MiniMax H3; for every other model, nobody has published one we could find.
What the card is
| Spec | RTX 3060 |
|---|---|
| SpecMemory | RTX 306012 GB GDDR6 on a 192-bit bus. An 8 GB model exists, on 128-bit |
| SpecArchitecture | RTX 3060Ampere, third-generation Tensor Cores |
| SpecCUDA cores | RTX 30603,584 |
| SpecUsable VRAM | RTX 306012,288 MiB in total; about 11.6 GiB once a desktop is running, in one owner's reading |
Check which one you have. The 8 GB RTX 3060 is a different card for this purpose, and the name alone does not tell you.
Ampere has no FP8 Tensor Cores. The fp8 files that ComfyUI's repacks use still load on a 3060 and still halve the memory they take; they are not faster than bf16 on it. That point comes from LightX2V's quantisation documentation, as quoted by smeltcore; we did not open the original.
Model by model
Wan 2.1 1.3B
Wan's README says the 1.3B text-to-video model "requires only 8.19 GB VRAM", for a 5-second 480p clip on an RTX 4090 in about 4 minutes. The efficiency table's note says that 4090 run used --offload_, so the text encoder was on the processor. The ComfyUI fp16 file is 2.84 GB. By file size, it is the easiest model on this page for a 3060.
Wan 2.2 5B
ComfyUI's tutorial says the 5B "should fit well on 8GB vram with the ComfyUI native offloading". Wan's own script asks for 24 GB at 720p. The two are not in conflict: they are different programs. The fp16 file is 10.00 GB; the full template set is 18.14 GB on disk. More on our Wan 2.2 VRAM page.
Wan 2.1 and Wan 2.2 14B
The template's fp8 files are 14.29 GB each, and Wan 2.2's 14B needs two of them, a high-noise and a low-noise expert. Neither fits on the card, so ComfyUI streams part of each from RAM.
GGUF builds are the usual 3060 route. The Q4_K_M of Wan 2.1 14B is 10.12 GB; each Wan 2.2 14B expert at Q4_K_M is 9.65 GB. Both fit by size, with about 2 GB left for working memory. They load through the ComfyUI-GGUF custom node.
Wan's script quotes 80 GB for the 14B at 720p. That is its own program, not ComfyUI.
HunyuanVideo 1.5
Tencent's README gives a minimum of 14 GB of VRAM, measured with model offloading on. ComfyUI's tutorial describes it as targeting 24 GB consumer cards. The ComfyUI repack's 480p fp8 model is 8.33 GB, and a Q4_K_M GGUF is 5.09 GB.
Its text encoder, Qwen 2.5 VL 7B, is 9.38 GB in fp8. With the fp8 model, the glyph encoder and the 2.52 GB VAE, the 480p set is 20.68 GB on disk.
LTX-2.5
Lightricks' minimum is 32 GB of VRAM, and its own guide calls anything below unsupported. The smallest official files are int8: 21.50 GB for the model and 15.37 GB for the text encoder.
One person published a measured run on a 12 GB card, an RTX 4070, not a 3060: 11.4 GB of VRAM and 45 GB of system RAM for a 5-second clip. The workflow, the GGUF options and the settings are on our LTX Video 12GB page.
CogVideoX 2B and 5B
The model cards give the lowest figures of any here: from 4 GB for the 2B and 5 GB for the 5B. Those are for the diffusers library with every offload and tiling option on, tested on A100 and H100 cards; the cards say peak use is about three times higher without them. The 5B transformer is 11.14 GB and its T5 text encoder 9.52 GB.
Mochi preview
Genmo's README says about 60 GB on one GPU in its own code, and that ComfyUI can run it in under 20 GB. The ComfyUI fp8 model is 10.03 GB. The all-in-one fp8 checkpoint is 15.84 GB, because it also carries the text encoder and VAE; the diffusion_models vs checkpoints page explains the difference.
MiniMax H3
This is the one we measured. On our RTX 3060 12GB, the official reference-to-video workflow made a 1344×768, 124-frame clip in about 43 minutes, twice. Both runs peaked at 11,649 MiB of VRAM; system memory peaked at 43,587 and 43,907 MiB. A run about 40 times smaller cut the time to under a minute but barely moved the memory. The full test card has the method, hashes and versions.
One other owner publishes measured RTX 3060 runs: the runs-on-a-3060 repository on GitHub. Its MiniMax H3 log, at 864×480 and 243 frames with a 4-step Turbo LoRA, took 6 to 7 minutes per clip on a machine with 64 GB of RAM, and it names system RAM, not VRAM, as the binding limit.
System RAM is the spec to check first
Every model above that does not fit in 12 GB runs by keeping the rest in system RAM. The measured figures agree:
| Run | Peak VRAM | Peak system RAM |
|---|---|---|
| RunMiniMax H3, our RTX 3060 | Peak VRAM11,649 MiB | Peak system RAM43,587 MiB |
| RunLTX-2.5, an RTX 4070 12GB | Peak VRAM11.4 GB | Peak system RAM45 GB |
A 12 GB LTX workflow on CivitAI asks for 48 GB of RAM, and Wan's 14B template keeps one 14.29 GB expert and a 6.74 GB text encoder in RAM while the other expert works. If you have 16 GB, the small models above are your range. With 32 GB, the 14B models in GGUF become possible. For the largest ones, plan on 48 to 64 GB.
That last paragraph is our reading of the figures above, not a measurement of each model.
Why the numbers online disagree
| Source | What it says for a 3060 | Measured? |
|---|---|---|
| SourceModel READMEs | What it says for a 3060Figures for their own scripts, on 4090, A100 or H100 cards | Measured?Yes, but not on a 3060 and not in ComfyUI |
| SourceWill It Run AI | What it says for a 3060Per-model VRAM at each model's full resolution, e.g. about 21.6 GB for Wan 2.1 1.3B | Measured?No. The site says its figures are estimates; they leave out offloading, which is why it is far above Wan's 8.19 GB |
| Sourceheiss-ui | What it says for a 3060"Fits" or "Offloads" per model | Measured?No for video. The verdicts follow file sizes; its measured timings are for image models |
| Sourcesmeltcore | What it says for a 3060A LightX2V recipe for Wan 2.1 14B on a 3060, INT8 with offload | Measured?No. It says no RTX 3060 benchmark has been published and leaves the speed out |
| Sourcevgoodslab | What it says for a 3060LTX-2 19B GGUF "tested and confirmed working on a 3060 12GB" | Measured?No VRAM reading or timing |
| Sourceruns-on-a-3060 on GitHub | What it says for a 3060MiniMax H3 at 864×480 | Measured?Yes |
| SourceOur test card | What it says for a 3060MiniMax H3 at 1344×768 | Measured?Yes |
Wan2GP, the other route
Wan2GP is a separate app built for low-VRAM cards. Its README says select models run "with as little as 6 GB of VRAM" and supports cards back to the GTX 10 series. Its memory profiles trade VRAM for RAM: profile 4, the default, streams the main models part by part; profile 5 streams almost everything, for machines short of both. These are the developer's own figures, with no stated test hardware.
What nobody has published
- A measured peak VRAM on an RTX 3060 for Wan, HunyuanVideo 1.5, LTX, CogVideoX or Mochi in ComfyUI.
- The same model measured at two resolutions on one 3060, to show how much the working memory moves.
If you own one and run these, record peak VRAM, peak system RAM and time, the way our test card does.
To check a card and RAM size against MiniMax H3's presets, use the system requirements checker. The other models on this page are not presets yet.
We do not host any of these models. Download them from the repositories named below, and read each model's licence on its own page.
GenVidKit is an independent guide. It is not affiliated with NVIDIA, MiniMax, Alibaba, the Wan team, Tencent, Lightricks, Zhipu AI, Genmo, Comfy Org, the Wan2GP developers or Hugging Face.
Sources
All read on 2026-10-09.
- NVIDIA RTX 3060 specifications — memory, bus, cores.
- Wan2.1 README and Wan2.2 README — Wan's own VRAM statements.
- ComfyUI Wan 2.2 tutorial and HunyuanVideo 1.5 tutorial — ComfyUI's statements and files.
- HunyuanVideo-1.5 README — the 14 GB minimum with offloading.
- ComfyUI-LTXVideo and Lightricks/LTX-2.5 — LTX requirements and files.
- zai-org/CogVideoX-2b and CogVideoX-5b — CogVideoX figures and files.
- Mochi README and Comfy-Org/mochi_preview_repackaged — Mochi figures and files.
- Hugging Face file lists: Comfy-Org/Wan_2.1_ComfyUI_repackaged, Comfy-Org/Wan_2.2_ComfyUI_Repackaged, city96/Wan2.1-T2V-14B-gguf, QuantStack/Wan2.2-T2V-A14B-GGUF, Comfy-Org/HunyuanVideo_1.5_repackaged, jayn7/HunyuanVideo-1.5_T2V_480p-GGUF.
- Wan2GP README and CLI documentation — the low-VRAM claims and profiles.
- runs-on-a-3060 — the other measured 3060 log.
- Will It Run AI, heiss-ui, smeltcore and vgoodslab — the third-party pages compared above.