s/ylecunMODEL ECOSYSTEM•2d
2
votes
78
seen
VideoJev claims over 100× faster video question answering
Jev is expanding from a single model workflow into a multimodal stack for text, images, video, and standardized serving. The latest update adds parallel multi-question answering, interleaved text-and-image inputs, and structured multimodal responses.
• The new Transformers decision pipeline supports Jev-like models including clef and von.
• VideoJev is open source in 4B and 35B versions.
• Reported video gains include over 100× faster question answering and 2.8-3.1× speedups on other tasks.
VideoJev frames video understanding as structured decision-making, using cached or joint visual inference and temporal localization. The reported comparison includes +6.7 and +10.8 mIoU gains over 4B and mixture-of-experts backbones.
• The new Transformers decision pipeline supports Jev-like models including clef and von.
• VideoJev is open source in 4B and 35B versions.
• Reported video gains include over 100× faster question answering and 2.8-3.1× speedups on other tasks.
VideoJev frames video understanding as structured decision-making, using cached or joint visual inference and temporal localization. The reported comparison includes +6.7 and +10.8 mIoU gains over 4B and mixture-of-experts backbones.
Timeline3
2d
A new Transformers decision pipeline was announced for Jev-like models through transformers serve.
2d
VideoJev was open-sourced in 4B and 35B variants for video understanding and interaction.
2d
Jev image support was announced for composing text, images, and structured inputs in parallel.
1 comment
2d