Asking a video model to answer questions today means generation: tokens stream out one by one, and thinking modes prepend a long chain of reasoning. For visual agents sitting inside an interaction loop, every "what should I do next" decision waits on a full generation pass, and that latency compounds. VideoJev, released by the MCG-NJU team at Nanjing University, takes a different route: it casts video understanding as structured decision tasks where the model scores instead of speaks. The paper hit #1 on Hugging Face Daily Papers the same day (175 upvotes), with training/evaluation code and two open checkpoints (arXiv:2610.18615).

Score, don't generate

The method surgically adapts pretrained multimodal LLMs: keep the vision and language backbones, replace autoregressive answer decoding with a shared head that scores temporal intervals and textual options against video evidence. The scores yield event spans for temporal grounding or answer choices for multiple-choice QA — two task families, one decision interface. This extends the System One "fast thinking" paradigm that has been heating up this year: Microsoft Decision-1, Cloudflare Clef, and a GLM-5.3-Flash-based approach already validated scoring-over-generation on text tasks; VideoJev moves the playbook to video. The team built VideoJev-4B and VideoJev-35B-A3B on Qwen3.5-4B and Qwen3.6-35B-A3B backbones.

More accurate, and absurdly faster

Across three temporal grounding benchmarks, VideoJev-4B and 35B-A3B improve mean mIoU over their respective Qwen backbones by 6.7 and 10.8 points, with measured inference speedups of 2.8-3.1x — accuracy and speed improve together, not traded off. Video QA is where it gets extreme: on four benchmarks, both models perform comparably to their backbones' Thinking mode while running over 100x faster on model forward passes. For agents that repeatedly query the same video, this is an order-of-magnitude difference.

The other half of fast: visual computation reuse

Two inference modes power the speedup: Cached inference reuses already-computed visual representations when new queries arrive, skipping a full re-watch; Joint inference packs multiple questions into a single forward pass. As an interaction unfolds, new goals and intermediate results keep prompting queries against the same visual history — the reuse design exists precisely for that loop.

Open delivery

The repo ships under Apache 2.0 with 26 commits and 46 stars. Training runs two stages: Head warmup (scoring head only, BCE+CE losses), then a Joint stage that trains the language backbone together with the head (adding duration-weighted SoftIoU), with the vision encoder frozen throughout; each stage runs one epoch at global batch 64. Training data comes from the open TimeLens-100K and TimeLens2-93K datasets, the framework is XTuner, and evaluation fixes a 256K context. Both checkpoints are BF16 Safetensors; the 4B demo runs in about ten lines of code.

So what

Video is a heavier modality than text: every frame costs compute, and every generated token bills the agent that just wants to look and decide. VideoJev's signal is not any single benchmark number but the interface-level unification — grounding, QA, and multi-query reuse share one scoring mechanism that complex workflows can compose directly. Now that fast-thinking has stretched from text into video, the next question might be: which other modalities are waiting to stop generating and start scoring?