Is PrimeSim state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.
PrimeBot Research's action-conditioned, multiview robot world-model simulator, distributed through Qiyuan Robot's Hugging Face namespace and PrimebotTech's code repository. Given each camera's first frame, a prompt and a robot action trajectory, it jointly generates three synchronized views. It adapts the roughly 5B Wan2.2-TI2V-5B backbone through midtraining and posttraining. Apache-2.0 weights and training/inference code are public; no hosted API is confirmed.
The details so far
- Lab
- PrimeBot Research
- Announced
- Nov 3, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- ~5B transformer backbone
- Lab identity
- Paper/project affiliation: PrimeBot Research; Qiyuan Robot correspondence domain and qiyuan-robot HF namespace; code published as PrimebotTech/PrimeSim
- Inputs and outputs
- Each camera's first frame, text prompt and robot action trajectory produce three synchronized camera-view videos; not an action-generating robot policy
- Base
- Wan2.2-TI2V-5B; midtraining adds action conditioning and multiview adaptation, followed by causal few-step posttraining
- Training corpus
- Lab reports 30,110 hours of cross-embodiment demonstrations across 22 robot platforms, mixing real and simulated data; corpus is not distributed
- Checkpoints
- primesim_base.safetensors: midtrained model for adaptation. primesim_posttrain_g2.safetensors: Stage-4 causal few-step model for inference
- License
- Apache-2.0 for released weights (HF card) and code (repository LICENSE); training corpus is excluded
- Requirements
- Weights are not self-contained: Wan2.2 text encoder, tokenizer and VAE are required. HF recommends an A100/H100 80 GB; combined resident transformer and text encoder do not fit a 24 GB card
- Limitations
- New-platform adaptation requires that platform's data and gradient updates, not zero-shot transfer. Closed-loop policy improvement is validated in simulation only, not physical hardware. Speed and quality evaluations use different resolutions and samplers
- Related
- Separate from Wan2.2-TI2V-5B and unrelated robot policies/world models; base and posttrain_g2 are checkpoints of one PrimeSim model
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: approximately 10.1 temporal FPS (10.12 rollout steps/s, 30.35 view-frames/s across three views) on one NVIDIA A100-80GB, BF16, batch size 1, four-step Stage-3 sampler at 288 x 384 per view. Short-window steady-state timing excludes loading/warm-up; longer continuous-KV episodes measure 7.63 temporal FPS. This is not end-to-end policy/control-loop speed or the 192 x 256, 50-step quality-evaluation setting.
Lab claim: action-following rollouts after 100–200 gradient steps on held-out xArm 7, ViperX 300 and UMI gripper, using their own data at batch size 1 and learning rate 3e-5; pretrained Wan2.2 baseline did not reach that behavior within 2,000 steps. This is few-shot adaptation, not zero-shot.
Sources
LABPrimeSim paper v1 and public-date metadataarxiv.org/abs/2611.04870v1 · Nov 3, 2026LABPrimeSim paper: throughput protocol and limitationsarxiv.org/html/2611.04870v1LABPrimeBot Research PrimeSim projectprimebottech.github.io/PrimeSim-website/REPOQiyuan Robot PrimeSim model card and checkpointshuggingface.co/qiyuan-robot/PrimeSimREPOPrimebotTech PrimeSim codegithub.com/PrimebotTech/PrimeSimREPOPrimeSim code Apache-2.0 LICENSEgithub.com/PrimebotTech/PrimeSim/blob/main/LICENSETracking continues until an independent board lists it.