LongLive2.0-Robot-L Robot Video Generation Model: File List

LongLive2.0-Robot-L Robot Video Generation Model: File List

At a Glance

Item Value
Repository Efficient-Large-Model/LongLive2.0-Robot-L
Publisher guide Efficient Large Model (NVIDIA and MIT): models and licenses
Published 2026-10-28
License other
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Efficient-Large-Model has released the weights for “LongLive 2.0 Robot-L," a model for generating robot videos. This model is an autoregressive image-to-video model that takes an image as input and outputs a video of robot motion.

This model is the final checkpoint of “Stage-L," which underwent continuous training on longer sequences following Robot-M, and includes the weights at step 3,972. Note that the “S," “M," and “L" designations in the model names do not refer to parameter size, but rather to the differences in sequence length during training.

Specifications

The specifications listed in the documentation are as follows:

  • Architecture: autoregressive
  • Parameter count: 4,999,787,712
  • Generation type: image-to-video
  • Weight details: 825 FP32 tensors (optimizer, EMA, and other training states are not included)
  • License: other (the terms of the original upstream model apply)

The weights in this release only include the online generator and are not a standalone pipeline. Therefore, to perform inference, you must separately prepare a compatible inference runtime, VAE, and text encoder.

Performance

Regarding model validation, the following has been reported:

All 20 Stage-L fixed-video validation gates were approved. The final evaluation panel includes 24 videos evaluated with settings of 50 sampling steps, CFG 3.0, and base seed 20260711.

However, it is explicitly stated that these validations are presented as release validation and do not guarantee safety in robot control or performance on specific benchmarks.

Strengths and Use Cases

This model is primarily intended for video generation in the field of robotics. It employs an autoregressive architecture and specializes in the “image-to-video" task of generating robot motion videos from images.

The “L" in the model name refers to the sequence length used in training. This model is the final checkpoint of “Stage-L," which underwent continuous training on longer sequences following Robot-M. As such, it excels at generating robot motions over longer time axes.

Distributed Files

Weight files published in Efficient-Large-Model/LongLive2.0-Robot-L, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
model.pt 20.00GB

How to Get It

The model weights can be downloaded from the Hugging Face repository using the following command:

hf download Efficient-Large-Model/LongLive2.0-Robot-L --local-dir./weights/LongLive2.0-Robot-L

Several preparations are required when performing inference, as the weights in this release only include the online generator and are not a standalone pipeline. Specifically, you need to separately prepare a compatible LongLive 2.0 generator, VAE, and text encoder.

For the compatible inference runtime and specific inference procedures, please refer to the following guide: Long-WAM Stage 1 guide

Use the following PyTorch code to load the weights:

import torch

state_dict = torch.load("model.pt", map_location="cpu", mmap=True, weights_only=True)
model.load_state_dict(state_dict, strict=True)

Can You Run It Locally?

License — other: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (other). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Related Articles

What to Read Next

Sources