PrimeSim

An action-conditioned, multiview world model that runs as a robot simulator. Given the first frame of each camera view, a prompt and an action trajectory, it generates what happens next.

• Paper: arXiv 2611.04870
• Project page: primebottech.github.io/PrimeSim-website
• Code: github.com/PrimebotTech/PrimeSim -- inference and training code

Trained on 30,110 hours of cross-embodiment data spanning 22 platforms. In simulation it leads four of five WorldSimProbe suites and posts the best overall EWMBench score.

Files

file	use
primesim_base.safetensors	the midtrained model -- start here to fine-tune
primesim_posttrain_g2.safetensors	stage 4, causal few-step model -- use for inference

Run primesim_posttrain_g2 on the demos in the repository to see what posttraining adds.

Requirements

These weights are not self-contained. The text encoder, VAE and tokenizer come from Wan-AI/Wan2.2-TI2V-5B:

hf download Wan-AI/Wan2.2-TI2V-5B \
  --include "models_t5_umt5-xxl-enc-bf16.pth" "Wan2.2_VAE.pth" "google/umt5-xxl/*" \
  --local-dir checkpoints

Point config/model/primesim.yaml at the three resulting paths.

Inference

Three camera views are generated together at 192x256, running above 10 FPS. An A100 or H100 (80 GB) is recommended: the 5B transformer and the 11 GB text encoder are resident at the same time, which does not fit on a 24 GB card.

License

Apache 2.0. The training corpus is not distributed and is not covered by this licence -- the weights are released, the data is not.