VideoJev scores video answers directly, reporting 100x faster QA
The Nanjing University-led project released 4B and 35B-A3B models; its speed tests measure model-forward time on H200 GPUs, not end-to-end response time.
By Ryan Merket · Published
Primary source: X
Why it matters
VideoJev tests whether video agents can answer repeated, structured questions by scoring choices and reusing video computation instead of generating text each time. Its speed claims are limited to model-forward benchmark timing, leaving end-to-end deployment costs untested.
Researchers led by Nanjing University have released VideoJev, a pair of video-understanding models that score candidate answers and event times directly instead of generating a written response. In an October 14th post on X, Aran Komatsuzaki (@arankomatsuzaki) highlighted the project's claim of more than 100x faster video question answering and released 4B and 35B-A3B variants.