🖥️ Open-Weight AI Video Models You Can Actually Run
MiniMax H3 going open-weight restarts this argument. What runs on 8GB, what needs 24GB, and what the quantization actually costs you in quality.
Derek Holt · Local AI & Hardware Writer
· 4 min read
MiniMax announced frontier-grade video weights as open on July 31, and ByteDance pointedly did not. That makes this a good moment to take stock of what you can actually pull down and run, because the answer changed twice this year and most of the guides floating around are quoting VRAM numbers from a model generation that no longer exists.
Everything below is what fits on a real card, not what fits on a vendor's A100.
By the numbers
| Model | Practical VRAM | Notes |
|---|---|---|
| LTX Video | 8 GB+ | Fastest iteration; LTX-2.3 does native audio-video in one pass |
| Wan 2.2 (1.3B) | ~8 GB | The 12GB-card entry point |
| Wan 2.2 (14B) | 16–24 GB | Quality tier; runs on 24GB with INT8/FP8 |
| HunyuanVideo (slimmed 8.3B) | ~14 GB with offloading | Tencent cut it down from the original 13B |
| HunyuanVideo (full) | 24 GB | 47GB at native precision before optimization |
| MiniMax H3 | Unknown | Open weights rolling out; no credible figures yet |
That last row of the chart is the real story of the past year. FP8 weights, GGUF quantization and tiled decoding took HunyuanVideo from 47GB to roughly 8GB. The model did not get worse by a factor of six.
The 8GB tier is real now, with a catch
Two years ago local video meant a 24GB card or nothing. That is no longer true. LTX Video and Wan 2.2's 1.3B variant both run in 8GB, which puts a 3060 Ti or a 4060 in the game.
The catch is that 8GB buys you the small models, not small versions of the big ones. Wan 2.2 14B quantized down to fit in 8GB is not the same thing as Wan 2.2 14B — you lose the parameter count that produced the quality you wanted. The honest framing is that 8GB gets you fast iteration and draft-quality motion, and you finish elsewhere.
On my 4080 the sweet spot has been Wan 2.2 14B at FP8. Sixteen gigabytes of the card's VRAM does the work, the rest gives the tiler room, and I am not fighting OOM errors on every second render.
Speed matters more than the spec sheet suggests
Raw VRAM tells you whether a model loads. It tells you nothing about whether you will actually use it, and that comes down to iteration time.
Community benchmarking on Wan I2V at 720p put a baseline render at 542 seconds and the same job with SageAttention plus TeaCache at 248 seconds. That is the difference between six renders in an hour and thirteen. On a model where you are re-rolling for motion you like, doubling your attempts per hour is worth more than a marginal quality bump.
If you have installed a local video model and left the speed stack alone, that is the first thing to fix — before buying a bigger card.
What LTX-2.3 does that the others don't
Native audio and video in a single pass, and it is the only open model doing it. That mirrors what MiniMax H3 does on the API side, and it matters for the same reason: a silent clip carries an hour of audio post behind it.
It is not as strong as Wan 2.2 14B on pure picture quality. For dialogue-adjacent work where the sound has to land on the frame, the tradeoff usually goes LTX's way anyway.
Where H3 lands, once it lands
If the H3 checkpoint arrives as promised, it becomes the first genuinely frontier-tier video model you can point at your own hardware — 2K, native audio, and reference control that the current open crop does not match. Read the full H3 breakdown for what the model does.
Temper expectations on hardware. A 2K-native model with an audio head is not going to be kind to a 12GB card, and the first week of community quants is always the worst week of community quants. I would not plan a build around it until somebody has posted real numbers on a consumer card.
How we tested
Wan 2.2 and the speed-stack figures come from our own rig — an RTX 4080, FP8 weights, SageAttention and TeaCache — cross-checked against the Civitai community benchmark on the same workload. HunyuanVideo and LTX VRAM ranges are from published optimization guides and the model cards, not from a vendor deck.
The H3 row is deliberately empty. Weights were still rolling out as of August 1 and we do not publish numbers we have not seen.
What didn't make the cut
CogVideoX and Mochi. Both still run, both are behind the current tier on motion quality, and neither is where I would send someone starting today.
We also left out multi-GPU configurations. They work, they are how you would serve this at volume, and they are not what somebody with one desktop and one graphics card is asking about.
Pros and cons
Worth going local for: client footage that cannot leave your machine, high iteration counts where per-second API billing punishes re-rolls, and permanence — a downloaded checkpoint cannot be repriced or retired out from under you.
Skip it for: one-off projects, 4K masters, and anything where the total render volume is under a few thousand seconds. At that scale the API is cheaper than your electricity plus the afternoon you will spend on CUDA versions.
Frequently asked questions
▸What's the minimum VRAM for local AI video generation?
8GB gets you in the door. Wan 2.2's 1.3B variant and LTX Video both run there, and a quantized HunyuanVideo will too with GGUF weights and tiling. What 8GB does not get you is the 14B-class quality tier — that wants 16-24GB, and no amount of offloading makes a 3060 render like a 4090.
▸Is MiniMax H3 available to download yet?
The weights were still rolling out as of August 1, 2026. MiniMax released H3 as open-weight while ByteDance kept Seedance 2.5 closed, but until the checkpoint is actually downloadable there are no credible consumer-GPU VRAM figures. Anything you see quoted this week is extrapolation.
▸Does quantization ruin video quality?
Less than you would expect on the picture, more than you would expect on motion. FP8 on a 14B model is close to free on a static-ish shot. Push to aggressive GGUF quants and the first thing that degrades is temporal consistency — flicker between frames and detail crawling on textures — before still-frame quality visibly drops.
▸Is running video models locally cheaper than an API?
Only past real volume. At $0.13 a second on a cloud model, a 24GB card pays for itself somewhere north of 2,000 seconds of finished video, and that ignores your electricity and your time. Local wins on privacy and on iteration count, not usually on a small project's invoice.
The 5 best AI video finds, every week
New models, tested prompts, and what actually worked in our production — one short email a week. No spam, unsubscribe anytime.
Written by Derek Holt
Local AI & Hardware Writer
Runs the site's local-inference rig and benchmarks every GPU, quant, and speed-stack claim on it personally before it goes in a guide. Will not shut up about VRAM bandwidth.
Keep learning
GuidesRun Wan 2.2 Locally: Free AI Video on Your Own GPU (2026)
Our exact local Wan 2.2 setup on an RTX 4080: GGUF quantization, the Triton + SageAttention + TeaCache speed stack, and the mistakes that waste a weekend.
2026-07-17
GuidesHunyuan 3 Local Guide: Can You Actually Run 295B at Home?
Tencent open-sourced Hunyuan 3 — 295B total, 21B active, Apache-2.0. We break down the real VRAM math, the GGUF quant path, and who should skip it entirely.
2026-07-31
GuidesMiniMax H3 Guide: 2K Video, Native Audio & Open Weights
MiniMax H3 landed July 31 with native 2K, built-in dialogue and open weights. The real API limits, the five input modes, and where it still breaks.
2026-08-01
GuidesBest GPU for Local AI in 2026: A VRAM-First Guide
The GPU guide written from a rig that renders AI daily: why VRAM beats speed, what each budget tier really runs, and the used cards that embarrass new ones.
2026-07-10
ComparisonsLocal AI vs Cloud APIs: The Real Cost Math (2026)
We run both daily. When local hardware beats per-token APIs, when cloud wins, and the break-even math nobody shows — electricity, amortization and rental included.
2026-07-17