LongCat Distilled, Explained
The faster, cheaper LongCat-Video checkpoint — what it is, what it costs, and when to pay a little more for the standard model.
Short answer: LongCat Distilled is the distilled (lighter, faster) checkpoint of the LongCat-Video model. On this site it runs at 480p or 720p for fewer credits than the standard model — Distilled 480p is 150 credits, Distilled 720p is 250 credits, versus 200 and 300 for the standard model — in exchange for a shorter clip and slightly lower detail. The catch is a hard cap: distilled runs stop at 81 frames (about 5.4 seconds at 15 fps), and asking for more silently switches the job to the standard model at its higher price.
What "distilled" actually means
Distillation is a training technique that makes a smaller, faster model by having it learn from a larger "teacher" model's output. The distilled model does the same job in fewer sampling steps, so it renders faster and costs less to serve — but it trades away a little quality to get there.
For LongCat specifically, the picture splits in two:
- The hosted LongCat-Video "Distilled" tier. This site and its upstream provider serve two checkpoints of the LongCat-Video model — a standard one and a distilled one. The distilled checkpoint uses fewer sampling steps, which is why it is faster and cheaper. The site's own guidance is direct: "720p looks cleaner, 480p is faster, Distilled 480p is the quickest." The specific step count of the hosted distilled checkpoint is not stated on the model's public page.
- The open model's use of distillation. Meituan's LongCat-Video repository uses step distillation explicitly in LongCat-Video-Avatar 1.5, which "accelerates inference to 8 steps via step distillation" and requires a
--use_distillflag to run at all. That is the same idea applied to the audio-driven avatar model, and it is the clearest official statement of what "distilled" means in the LongCat family.
LongCat-Video itself is Meituan's open-weights video model: 13.6B parameters, MIT-licensed, and built for text-to-video, image-to-video and video continuation in one model. (Source: Meituan LongCat-Video repository, verified 2026-10-21.)
The four options and what each costs
The generator offers four LongCat-Video configurations. Cost is flat per generation — the frame count you request does not change the price:
| Option | Resolution / motion | Frame range | Credits |
|---|---|---|---|
| Distilled 480p | 480p, fewer sampling steps | 17–81 frames | 150 |
| Standard 480p | 480p / 15 fps preset | 17–961 frames | 200 |
| Distilled 720p | 720p, fewer sampling steps | 17–81 frames | 250 |
| Standard 720p | 720p / 30 fps preset | 17–961 frames | 300 |
Two details trip people up:
- Distilled is capped at 81 frames. Request more and the generator upgrades the run to the standard model automatically, and the quote changes with it.
- The price does not grow with length. A 162-frame and a 961-frame standard 480p job both cost 200 credits; longer jobs just take longer to finish.
The full frame-to-seconds math is in How Long Can LongCat Video Generate?.
How to use the distilled model
- Open the LongCat Video generator and pick text-to-video or image-to-video.
- In the video settings, choose a Distilled tier — 480p for the fastest and cheapest result, 720p for a cleaner fast clip.
- Keep the frame count at or under 81. The moment you ask for more, you are billed as the standard model.
- Write a prompt with one subject and one camera move, generate, and read the motion before committing more credits.
- For the final render, switch to a standard tier and raise the frame count.
Use distilled runs as the cheap prototype step: at 150 credits it is the least expensive way to confirm framing, subject and motion are right before a longer standard render. See the ultimate guide for prompt and parameter depth.
When to pick distilled vs standard
Pick distilled when:
- You are iterating on a prompt and want feedback in the fewest minutes.
- The clip is short by design — a hook, a loop, or a single action under 5 seconds.
- You are new to the tool and want to learn the controls at the lowest per-run cost.
Pick standard when:
- The shot needs to run longer than 81 frames (the distilled ceiling).
- You want maximum detail and smoother motion for a final render.
- Identity, product or text must hold up across a longer, more complex shot.
As a rule: prototype distilled, deliver standard. Start from an image when subject or product identity matters — the image-to-video mode anchors the first frame so the distilled run has less to invent.
What distilled costs against the standard model
On this site the distilled tier is always cheaper than the standard tier at the same resolution — 150 vs 200 at 480p, and 250 vs 300 at 720p. The same pattern holds on the upstream provider's pricing: the distilled 480p text-to-video endpoint is billed at $0.005 per generated second, against $0.025 for the standard 480p endpoint (source: fal.ai model pages, verified 2026-10-21). The site's credit system is its own; see pricing for the live credit packs.
Frequently Asked Questions
What is LongCat Distilled?
The distilled checkpoint of the LongCat-Video model: a lighter, faster version that renders in fewer sampling steps. On this site it costs 150 credits at 480p and 250 at 720p, and it is capped at 81 frames per generation.
Is LongCat Distilled better than the standard model?
It depends on the job. Distilled is better for speed, cost and short clips; standard is better for length (up to 961 frames) and maximum detail. Prototype with distilled, then render the final take with the standard model.
Why did my distilled job cost more than I expected?
Distilled runs are capped at 81 frames. If you request more, the generator automatically uses the standard model and applies its higher price, so the quote changes with it.
How long can a distilled LongCat video be?
Up to 81 frames. At 15 fps that is about 5.4 seconds, about 3.4 seconds at 24 fps, and about 2.7 seconds at 30 fps. The standard model goes up to 961 frames.
What does "distilled" mean for an AI model?
Distillation trains a smaller, faster model to imitate a larger teacher model, typically with fewer sampling steps. Meituan’s LongCat-Video repository documents this approach in LongCat-Video-Avatar 1.5, which reaches 8 inference steps via step distillation.
Prototype Distilled, Deliver Standard
Run cheap distilled clips to check the motion, then switch to the standard tier for the final render — no local GPU, no install.
Related guides
- How Long Can LongCat Video Generate? — frames, fps and the 81-frame distilled cap in detail.
- LongCat Video Ultimate Guide 2026 — distilled vs regular checkpoints, prompts and credit costs.
- LongCat FP8 workflow — how the Distill LoRA speeds up the open model in ComfyUI.
- LongCat Video pricing — current credit packs and per-job costs.
Sources: Meituan LongCat-Video repository (verified 2026-10-21); fal.ai LongCat Video Distilled and LongCat Video model pages (verified 2026-10-21). Credit costs and the 81-frame cap mirror this site's own pricing and generator configuration.