14B checkpoint

FastFrames-L: the 14B model and what it needs

About 28 GB of weights in FP16, before the encoder, the decoder and a single frame of activation. Here is what that means for your card, and the four levers that make it fit.

FastFrames-L is the 14B checkpoint. It is the one people screenshot and the one that breaks a casual setup, so it deserves a page that is honest about the trade.

What 14B actually costs

Start with the only number that does not depend on settings: the weights. Fourteen billion parameters stored in FP16 come to about 28 GB on disk and in memory. That is before the text encoder, before the decoder and before any activations for the frames you are sampling.

This single number explains most of the advice on this page. If your GPU has 24 GB, the weights do not fit at full precision and no flag fixes that — you have to change the precision or move part of the model somewhere else.

Four ways to make it fit

  1. Lower the precision of the weights. FP8 roughly halves the footprint and INT4-style quantisation cuts it further again. Quality moves; how much is something you should judge on your own prompt rather than trust a stranger's screenshot on.
  2. Offload to CPU or disk. Parts of the model live in system RAM and stream in as needed. It works, and it is slow — expect generation time to be dominated by transfers rather than by the GPU.
  3. Shorten the clip and drop the resolution. Memory scales with frame count and with the square of resolution. Halving resolution is worth far more than halving seconds.
  4. Use sequential or layer-by-layer loading if the pipeline you use supports it. Slower per step, but it turns a hard failure into a working run.

Which combination is available depends on the pipeline the model card documents. Read it there — the exact class names change between releases, and a snippet copied from a blog post is the most common reason a first run fails.

Check your disk before you check your GPU

A 28 GB download plus a cache copy is easy to underestimate. Keep at least 60 GB free, and make sure the Hugging Face cache directory is on the fast disk rather than a network mount. Half-finished downloads that resume badly waste more time than any tuning step.

When the large model is worth it

The honest answer: when quality is the product. Longer motion, cleaner detail and fewer collapse artefacts are what 14B buys you, and those differences show up most in the frames you would otherwise cut. For thumbnails, B-roll placeholders, storyboard tests and prompt iteration, FastFrames-S is the better use of your evening — see the 1.3B guide.

A practical split that works well: iterate every prompt idea on the small model, keep a shortlist, and render only the winners on the large one. You spend the expensive minutes on the clips that survive.

A realistic first session

If you are starting from nothing, do this in order. Try a community demo to see what the model is like before you commit a machine — the where to try page lists the live ones. Then pull the small checkpoint and get one clip out end to end. Only after that download the 14B weights, and only if you have both the disk and a plan for the precision. The local setup guide walks the environment and the usual errors.

One warning worth repeating: every memory number here is arithmetic from the parameter count, not a measurement of this specific release. Your real ceiling depends on resolution, frame count, sampler steps and the pipeline you use, so calibrate the first run before you promise anyone a delivery time.

Frequently asked

How much VRAM does FastFrames-L need?

The 14B weights alone come to roughly 28 GB at FP16, before the text encoder, decoder and activations. On a 24 GB card you cannot hold full precision weights, so you need quantisation, offloading, or a shorter and smaller clip.

Can I run FastFrames-L on a 24 GB GPU?

Yes, with changes: FP8 or INT4 weights, CPU offloading, or a combination. Expect generation to be slow and clip length to be limited. Without one of those, the run fails on the first allocation.

How big is the FastFrames-L download?

Roughly 28 GB of FP16 weights, plus the tokenizer, encoder and cache copies. Budget at least 60 GB of free disk and keep the Hugging Face cache on a local fast disk.

Is FastFrames-L better than FastFrames-S?

It produces stronger motion and finer detail, which is worth it for final renders. For prompt iteration, storyboard tests and short social clips, the 1.3B model is far cheaper in time and gives you more attempts per hour.