The Box launch gave Jev's typed answers a way to point at a still image. The round after it decides from whole clips, audio, and screenshots — work that has mostly landed in community checkpoints rather than the hosted API. Jev-Omni is the most complete, VideoJev handles video, and Laya Vision shows how small a decider gets. The contract is the same everywhere: media in, a probability distribution out.
One contract, four modalities
None of these models generate an answer. They are non-autoregressive deciders: one forward pass over the state, with a decision head scoring every declared option, so the response is a per-option probability. The earlier round on multimodal decision models covered the cloud APIs; these checkpoints are what you can run yourself.
Jev-Omni is the broadest of the batch: a merged classifier on Gemma 4 12B IT returning a probability for supplied text, image, audio, or video options. The card reports DecisionBench Medium 87.57%, JevBench 86.15%, MMAU 63.10%, and MVBench 53.10% — the author's own merged-model runs, with dataset rights separate from the Apache-2.0 weights. Warm H200 latencies: 83 ms for about 2,000 text tokens, 26 ms for an image, 31 ms for 13 seconds of audio, and 504 ms for a 16-frame video.
Caveats: a 50 GB FP32 footprint needing a CUDA GPU, a head that accepts 256 options while quality above 20 is not established, and no affiliation with TypeSafe AI.
Three ways to run Jev-Omni
| Route | Hardware | Measured | The catch |
|---|---|---|---|
| Replicate deployment | Nvidia L40S, hosted | About $0.00098 per run, typically under a second | No readme — inputs, formats, and output schema undocumented; the #1 Image JevBench v0.1 claim has no table behind it |
| Q4_K_M GGUF | RTX 5080 Laptop GPU | 8,980 MiB VRAM text-only; 9,256 MiB after an image decision | Probabilities drift; the llama.cpp adapter is unvalidated on Metal and Vulkan; image, video, and audio routes passed only smoke checks |
| MLX 4-bit | Mac mini M4, 16 GB | About 7.0 GB peak Metal memory; 963 ms median text, 994 ms image | An independent conversion with a post-hoc temperature fit — ECE-10 worsened from 0.06261 to 0.06774 on odd rows |
Both local builds keep the vision tower and 256-way head at higher precision while quantizing the language weights — 4.95 bits per weight in the GGUF, 4-bit affine in MLX. Aggregates look strong: 624 of 624 Persian-benchmark questions for the GGUF, 87.88% micro over 231 public JevBench decisions for the MLX build. Per-request numbers move more — the GGUF matched the winning option on 3 of 4 published examples, max difference 0.210 — and the Persian comparison pits two differently prompted systems against each other. Argmax survives conversion; the number you threshold on does not, so re-fit a calibration on your own media.
Video is its own decision problem
VideoJev treats clips as clips, not stacks of frames. VideoJev-4B (Qwen3.5-4B) and VideoJev-35B-A3B (Qwen3.6-35B-A3B), both BF16, answer two question shapes: grounding returns per-interval sigmoid probabilities with a 0.25 default threshold; multiple-choice returns softmax probabilities plus a zero-based answer index. The accompanying paper reports mean mIoU gains of 6.7 and 10.8 points over the backbones at 2.8–3.1x speedups — self-reported numbers on an October 13 preprint that names no released weights, so read them as directional. The quick start runs a demo clip at fps=2.0, max_frames=2048, context=8192; training freezes the vision tower across two stages, one epoch each. Budget for the toolchain, not the weights: Python 3.12, PyTorch 2.8 with CUDA 12.8, FlashAttention 2, causal-conv1d, and a compiled grouped-GEMM extension through XTuner for the 35B. For per-tick loops, see the per-tick article.
The small end of the spectrum
Laya Vision makes the opposite bet: replace Laya's text encoder with SmolVLM-256M and the decider is 201M parameters, not 12B. The recommended checkpoint reports A-OKVQA 59.8%, ScienceQA 82.4%, and VQAv2 yes/no 71.4%, with a 41 ms median predict on an L4 — a figure its README flags for roughly 50% host-to-host variance. Answers are choice, noul, and score over an image plus optional text in one forward pass; options are cut to 48 tokens, and a truncated input sets a truncated field. It also plays games from pixels: 0.99 ViZDoom basic, 0.81 Freeway, 0.76 Acrobot, against 0 for random play. The lab limits: the VQAv2 figure comes from an image-based re-split and is not comparable to published results, score answers are meaningful only on the two rubric-trained checkpoints, and the repo calls itself experimental and unaffiliated with Convai.
Screenshots without sending screenshots
typesafe-computer-use is the cheapest end of this round, barely a vision model: the loop reads the screen with macOS Vision OCR plus the accessibility tree, then sends one TypeSafe request with up to three Choices to pick the next action, so screenshots never leave the machine. On the same screenshot: $0.0002 per decision at 0.13–0.38 s of model latency, versus $0.032 and 5.2 s for Claude Opus 5. It runs up to 100 steps and stops on done or none, confidence below 0.4, two no-ops, a step limit, or Ctrl-C. Only the main display is captured, and icon-only buttons in terminals, canvases, or Spotify reach neither OCR nor the accessibility tree.
What to copy from this round
- The answer shape is stable: everything here returns probabilities over declared options and generates no text, so consumers built for typed decisions carry over unchanged.
- Option capacity is the binding constraint: Jev-Omni's head takes 256 options but quality past 20 is unestablished, and Laya Vision cuts options at 48 tokens. Shortlist before you ask.
- The cheapest screenshot loop is not a vision model: OCR plus an accessibility tree is enough when the screen contains text.
The model side is community work with separate licenses and its own caveats — but it is where pixels, clips, and audio get decided today.
