AaronHuangWei/longlive2.0-robot-m

LongLive 2.0 Robot — Stage-M final online generator

Public release of the accepted Core7 Stage-M model at training step 11,155. Stage-M continues from the accepted Stage-S online generator; it is not the later Stage-L model.

Weights

• File: stage_m_final_online_generator.pt
• Format: bare PyTorch state_dict, online generator only, FP32
• Parameters: 4,999,787,712; tensors: 825
• Size: 19999499609 bytes
• SHA-256: 38371710932738bafab6c74bfee32f05acfcc2e19a1c1b016965d09f778cbe8f

Optimizer state, EMA weights, RNG state, error buffers and intermediate checkpoints are excluded. Every exported tensor was compared exactly against the accepted M11155 source. The source checkpoint's established full-hash proof is recorded in the provenance; this publication does not claim a repeated full-source hash.

Loading

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("AaronHuangWei/longlive2.0-robot-m", "stage_m_final_online_generator.pt")
state_dict = torch.load(path, map_location="cpu", mmap=True, weights_only=True)
model.load_state_dict(state_dict, strict=True)

Instantiate the matching LongLive 2.0 / Wan2.2-TI2V-5B generator architecture first. This is a generator-weight release, not a standalone Diffusers pipeline; matching inference code, VAE and text encoder are required separately. No LoRA or few-step distillation is included. The established fixed evaluation uses 50 sampling steps, CFG 3.0, seed 20260711 and online weights. All 20 Stage-M evaluation gates were accepted; the final fixed panel contains 18 videos. These checks are workflow acceptance, not a claim of external benchmark performance or robot-control safety.

Provenance and terms

See stage_m_final_online_generator.pt.provenance.json for source binding, tensor schema, exact-value verification and excluded fields. SHA256SUMS provides the published file checksum. This release follows the generator-only format of the Stage-S repository. Derived from NVIDIA LongLive 2.0; applicable upstream model terms still apply. The upstream model card is at https://huggingface.co/Efficient-Large-Model/LongLive-2.0-5B and links the NVIDIA Open Model License. No new Apache-2.0 relicense is asserted here.