V-JEPA 2
Open video world model; the action-conditioned V-JEPA 2-AC variant plans robot manipulation zero-shot.
About
V-JEPA 2 (June 2025) is Meta FAIR's big bet on the JEPA idea: learn the world by predicting in representation space instead of generating pixels. The 1.2B-parameter video encoder is pre-trained self-supervised on over a million hours of video with no action labels, then post-trained into V-JEPA 2-AC, an action-conditioned latent world model using just 62 hours of Droid robot video. Deployed zero-shot on Franka arms in two labs, V-JEPA 2-AC did pick-and-place on novel objects at 65–80% success — Meta's numbers, but on real hardware, which counts for something. Everything is public: code, weights, and demos at github.com/facebookresearch/vjepa2, MIT/Apache licensed, with Hugging Face Transformers support. V-JEPA 2.1 followed in March 2026. The caveat is maturity: it is a research world model, not a turnkey robot stack, and independent robot reproductions are still thin.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Meta-reported: V-JEPA 2-AC 65–80% zero-shot pick-and-place on Franka arms, novel objects, two labs
- Motion understanding / action anticipation SOTA on video benchmarks (company-reported); independent robot reproduction not established
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: Meta FAIR release (June 11, 2025) with public code, weights, and demos; MIT/Apache licensed; HF Transformers integration; V-JEPA 2-AC tested zero-shot on real Franka arms in two labs. Strengths: fully open, reproducible weights, real-hardware robot results. Docked for: (1) robot numbers are company-reported; (2) independent robot reproduction not established. License: MIT / Apache-2.0 (per repo license files).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Meta FAIR release 2025-06-11; 1.2B video encoder pre-trained on 1M+ hours of video, no action labels
- · V-JEPA 2-AC action-conditioned variant post-trained on 62h Droid robot video; 65-80% zero-shot pick-and-place on Franka arms (Meta-reported)
- · Public code + weights at facebookresearch/vjepa2; Hugging Face Transformers support; V-JEPA 2.1 followed Mar 2026