Jev takes text, and only text. That was a firm line through 2026, and it is the first thing the rivals attacked. Cloudflare's Clef-omni adds images, audio, and video, Liquid's d1 adds images, and both keep the same typed-answer shape. Whether that matters depends entirely on whether your states contain pixels.
What changed
The decision-model contract — a state plus typed questions, answers with probabilities — was text-only by design. A multimodal decision model keeps the contract and widens the state to include media: Clef-omni accepts text, images, wav or mp3 audio, and mp4 or webm video in one request, and d1 accepts text, images, or both.
The output does not change. You still get a Choice, a Score, or a Noul. The difference is that a question can now be about a frame, a clip, or a recording instead of a description someone wrote first.
Where this genuinely helps
The text-only constraint forced a lossy pre-step: describe the image, transcribe the audio, summarize the clip, then decide. Every one of those steps is a place where the judgment input degrades. Multimodal decision models remove the pre-step for tasks where the media is the state:
- Screening an image against a content policy.
- Judging whether a product photo matches a listing.
- Classifying a screen recording or a UI clip.
- Routing on a voicemail or a call recording.
The browser and computer-use article shows the workaround that text-only Jev required: agents sent structured page observations instead of screenshots. A multimodal model lets the decision look at the pixels directly.
Where it changes nothing
For the overwhelming majority of decisions — routing a ticket, scoring a document, classifying a message — the state is already text and stays text. Adding an image modality does not change the primitive design, the threshold logic, or the state discipline. It adds an input type, not a new programming model.
The costs the vendors underplay
Three caveats recur across the multimodal entries:
- Tokens. Images, audio, and video become input tokens, and input is what you pay for. A clip costs far more than the sentence that would have described it, so the pre-step you removed may have been the cheap one.
- Vendor claims. The latency and quality numbers for Clef and d1 come from their makers. The price-war article collects the rival benchmarks and their sources; the FOMA comparison interleaves vendor and third-party figures. Measure on your own task.
- Calibration. A new modality is a new calibration problem. A model that routes well on text may be uncalibrated on audio, and your thresholds do not transfer.
What to do if you are text-only today
If Jev is your model and your states are text, the multimodal question is narrower than it sounds. Check whether your pipeline already does the describe-transcribe-summarize step, and whether that step loses the information the decision needs. If it does not — if the description is as good as the image for your question — stay text-only; you are paying less. If it does, a multimodal decision model is the first thing that competes with the workaround rather than the description.
Multimodal input widens the state, not the contract. The decision shape is the same; what you can point a question at is not.