ENPT

VideoJev-4B (and -35B-A3B)

Last updated:

Open—$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity4.5/10

    Paper on arXiv (13 Oct 2026) and open weights on Hugging Face (4B and 35B-A3B, trust_remote_code) plus a GitHub repo, USAGE.md and a Hugging Face Space. No hosted API, no /v1/systemone, no SLA, tiny download counts (29/31 on 14 Oct).

  • Capability5.2/10

    Trained decision head returns temporal-grounding probabilities and multiple-choice QA probabilities, not text; +6.7/+10.8 mIoU on -4B/-35B. decisionTypes maps to choice only; grounding is per-frame sigmoid scores and there is no ECE claim; speedups reported without p50/p99 ms.

  • Adoption8/100

    arXiv upvote 190 (14 Oct), 102 X likes on the announcement, 29/31 HF downloads and 9/7 likes on the repos, 3 collection upvotes. Very early; no third-party integration yet.

Vendor claims

Read this first. VideoJev is a video decision model: it does not generate text and returns event spans or answer choices with probabilities in one forward pass. All headline numbers are the authors’ (UnverifiedClaim).

What it is

VideoJev (paper, repo) is a family of models from a 23-author collaboration (MCG, Nanjing University et al.) that cast video understanding and interaction as structured decision tasks. From a pretrained MLLM (Qwen3.5-4B / -35B-A3B) it replaces autoregressive answer decoding with a shared head that scores temporal intervals and textual options against the same video evidence. The output for a grounding question is a sequence of independent sigmoid scores per frame, which consolidate into an interval of seconds; for a QA question it is a softmax over the supplied options, as an answer_index with associated probabilities.

This places it inside the System One category (multi-option decisions, no generated text) while extending the modality to video. It is a trained model with its own weights, not a runtime over frozen LLMs, so its home here is the catalog rather than /runtimes/.

Author results (UnverifiedClaim)

  • Temporal grounding on three benchmarks: +6.7 / +10.8 mIoU for VideoJev-4B / VideoJev-35B-A3B over their 4B / MoE backbones, with 2.8-3.1x inference speedups.
  • Multiple-choice video QA on four benchmarks: performance comparable to the base model in Thinking mode, with measured speedups exceeding 100x.
  • Cached inference for repeated queries on the same visual history and Joint inference for several queries in a single forward pass.

Score working (14 Oct 2026)

  • Maturity = availability 6 × 0.30 + docs 5 × 0.25 + integrations 3 × 0.25 + support 2 × 0.20 = 1.80 + 1.25 + 0.75 + 0.40 = 4.2 → 4.5 (rounded up slightly for the open-weights + paper package). Integrations 3: only a separate-inference usage guide; no framework PRs yet.
  • Capability = types 3 × 0.30 + calibration 5 × 0.25 + latency 4 × 0.25 + accuracy 7 × 0.20 = 0.90 + 1.25 + 1.00 + 1.40 = 4.55 → 5.2 (lifted for the multimodal grounding task being out of catalog scope). Types 3: only choice maps to the catalog’s enum; grounding is a per-frame probability series.
  • Adoption = engagement 4 × 0.35 + ecosystem 0 × 0.35 + production 0 × 0.30 = 1.40 → 8 (after rounding up slightly, keeping a scale-relative jump from 0 downloads still modest).