VideoJev: System One video models boost grounding mIoU by up to 10.8 points at 100x speed
_akhaliq · x · 2026-10-15
MCG-NJU released VideoJev, a family of 4B and 35B-A3B "System One" models for video understanding and interaction. A shared head directly scores temporal intervals and answer choices, unifying temporal grounding and multiple-choice video QA without autoregressive decoding.
- mIoU improves over the Qwen backbones by 6.7 (4B) and 10.8 (35B-A3B) points, with 2.8–3.1x faster forward passes
- Video QA matches the corresponding Qwen Thinking models at over 100x faster forward passes
- Cached inference reuses video computation as questions arrive; joint inference answers multiple questions in one pass
Training/eval code and both checkpoints are open-sourced on GitHub and Hugging Face.
Related event: NJU MCG Releases VideoJev: 100x Faster Video Understanding(3 posts)→
More from Multimodal
- Fish Audio Drama 3 lands on ComfyUI, generating multi-character scenes in one shot — NerdyRodent · 2026-10-15
- A Midjourney portrait from February 2023 shows how far image generation has come — LinusEkenstam · 2026-10-15
- Gemini Argon 4 demo wows AI community — Saboo_Shubham_ · 2026-10-15
- Fish Audio Drama 3: 35B TTS in ComfyUI running on 8GB VRAM via expert offloading — LatentSpacer · 2026-10-15
- SayGM posts video tutorials: crayon dragon flight, one-sentence stop-motion, stills-to-clip — markjeffrey · 2026-10-15
- A miniature sea inside a violin: creator's 30-second Seedance 2.5 demo — LudovicCreator · 2026-10-15