Introduction

PrimebotTech just cleared the weights gate on PrimeSim. Same-day (2026-11-06) they pushed full inference/training code to PrimebotTech/PrimeSim (release code @ ~09:05 UTC) and published two Apache-2.0 checkpoints on Hugging Face at qiyuan-robot/PrimeSim: primesim_base.safetensors (midtrain) and primesim_posttrain_g2.safetensors (stage-4 few-step inference) — about 10.1 GB each. Project page: primebottech.github.io/PrimeSim-website. Paper: arXiv 2611.04870.

PrimeSim is an action-conditioned, multiview world model built on Wan2.2-TI2V-5B. Give it the first frame of each camera view, a language prompt, and a robot action trajectory — it generates what happens next as synchronized video. Trained on 30,110 hours across 22 robot platforms (humanoid / dual-arm / single-arm / UMI), it is aimed at robot simulation and policy evaluation, but the same stack is the Titus-lane pattern we already track for playable / sim world models: open weights, runnable inference, action → visual consequence.

What shipped

  • Open code — primesim Python package, scripts/inference.py, Hydra train launches (finetune → AR → consistency distillation → DMD → Self-Forcing++), example demos under example_data/
  • Open weights — HF qiyuan-robot/PrimeSim with midtrain base + G2 posttrain stage-4 checkpoint (use posttrain for inference demos)
  • Modality — image(+actions+text) → multiview video world model; three views generated together (card cites 192×256, project page also quotes 288×384 / view) at few-step / ~10 Hz class rates on a single A100
  • Open vs API — Apache-2.0 weights + code; training corpus is not distributed; no hosted API claimed in the drop
  • Dependencies — text encoder / VAE / tokenizer still come from Wan-AI/Wan2.2-TI2V-5B (~14 GB extra); not a single-file bundle

Concrete maker fact: clone the repo, uv pip install -e ., download Wan2.2 TI2V aux files + primesim_posttrain_g2.safetensors, then python scripts/inference.py --checkpoint … --demo-dir example_data/g2_example_1. README is blunt that a 24 GB card will not fit the 5B transformer + umT5 encoder together — plan on A100/H100 80 GB.

Why game / sim pipelines should care

We already cover interactive world models for level scouting (Frameboy/Atlas) and open keyboard/WASD world models. PrimeSim is the robotics-adjacent cousin with three maker-relevant edges:

  1. Action tokens are first-class — a unified 31-dim cross-embodiment layout (EE pose, gripper, torso, head, waist, base) so rollouts follow commanded trajectories instead of free-wheeling video
  2. Multiview consistency — three cameras in one latent with view embeddings; useful if you are prototyping multi-camera NPC / vehicle / robot sims
  3. Fine-tune path included — midtrain recipe + LeRobot v3 adapters ship in-repo, so you can point it at your own embodiment data instead of waiting for a closed API

This is not a mesh exporter or a Unity plugin. Treat it as a simulator prior for action-conditioned previews and policy-in-the-loop eval, then rebuild geometry in DCC/engine when you need collision and art.

Where to run it today

Note: the marketing site still shows a stale “Weights Pending” badge in places — ignore it; the HF siblings and GitHub README are the primary zero-day artifacts.

Same-day takeaway

Soft watch closed: stub README → real primesim package + two ~10 GB safetensors on the same calendar day. If you are building action-driven world sims (robotics, vehicle AI, multi-cam NPC labs), start from primesim_posttrain_g2, keep Wan2.2 aux weights local, and budget an 80 GB GPU. Keep Frameboy for browser flythroughs; keep PrimeSim when you need open weights and action control.

—Titus