Post
CommunityJev Media & Talks
Clef-omni, a faster Clef and a cheaper Clef-flash
Cloudflare's announcement of Clef-omni, an open-weight decision model that takes audio and video alongside text and images in one request, built on a Qwen3-Omni mixture-of-experts backbone, plus a faster Clef and a cheaper Clef-flash.
- Format
- Post
- Reading time
- 3 min
- Published by
- Community
- Author
- Cloudflare
- Last verified
- Oct 2026
- Platforms
- Cloudflare
- Capabilities
- Open source
Highlights
- Clef-omni takes wav or mp3 audio and mp4 or webm video alongside text and images in one request, with no transcribe-then-classify pipeline.
- Built on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, keeping the comprehension backbone and dropping text-to-speech parts.
- Training: frozen backbone weights, low-rank adapters, and label-smoothed cross-entropy loss with Brier score calibration.
- The post says text-only decisions return in about 130 ms at the median on Workers AI.
- Clef-flash was cut to cheaper than Jev, per the post, and Clef got faster in the same October 9, 2026 release.
Watch out
Grounded on the captured Cloudflare blog text dated October 9, 2026; the capture is truncated mid-sentence after the latency and pricing passage, so later benchmark tables and caveats were not reviewed. The curl example in the post is cut before its URL in the digest, so no snippet is reproduced here, and the speed and price statements are Cloudflare's own claims.
Questions about this entry
Short answers drawn from the linked project or page.
Audio and video. It takes text, images, wav or mp3 audio, and mp4 or webm video in a single request, where earlier Clef supported images and video frame arrays; the post says this removes cascading pipelines that transcribe speech or split channels out of video.
A Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. Cloudflare keeps the primary comprehension backbone and discards the text-to-speech output components, then trains with frozen backbone weights, low-rank adapters and label-smoothed cross-entropy loss with Brier score calibration.
Clef models are not LLMs: a quick prefill pass scores all modalities and valid parameter options at once, and two-stage attention routing gathers evidence from text, visuals and audio before field vectors cross-attend to compute confidence. A built-in lexical grammar preserves option semantics.
The post says text-only decisions return in about 130 ms at the median, and that Clef-flash's price was cut so it is now cheaper than Jev. Those are Cloudflare's numbers from the announcement, and the captured text ends before any benchmark tables.
October 9, 2026, a week after Clef and Clef-flash. The post says the team decided to work on the decision-model space on a Friday evening, trained the model over the weekend, and launched on the Thursday.
The post shows a curl-style POST whose JSON carries model clef-omni, a state, images, audio and videos arrays with base64 data URIs, and a questions object of noul questions such as whether a label is visible or a fan is running. Authentication uses a Cloudflare token; the captured command is truncated before its URL.
No, Jev has been text-only since its debut, which the post marks as the state Clef-omni changes for decision models. If you need Jev on media today, you must transcribe or describe it first and pay for that extra step.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Search and tagging on keep.md
Reddit: Jev is on Workers AI as typesafe/jev
More Jev resources like this
RuntimeWire's report on Cloudflare extending its Clef decision models to audio and video: Clef-omni returns structured decisions from text, images, audio and video in one API call, while Clef-flash gets cheaper and its hosted context window shrinks.
Jev Media & TalksArticle#community#catalog-import#yibie
Community
The Apache-2.0 weights behind Cloudflare's Clef: a 27B multimodal model that takes text, JSON, images or video as state and returns a probability for every allowed option of every typed question in a single forward pass. The card links the Cloudflare announcement and a Decision Index leaderboard.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
Microsoft's announcement of Microsoft-Decision-1, a model for fast decision scoring built for routing, classification, prioritization, verification and workflow control, available in Microsoft Foundry and through OpenRouter. Published 2026.10.09.
Jev Media & TalksPost#community#catalog-import#madewithjev
Community
An open, local attempt to bring Jev-style typed decisions to images: a Qwen3-VL backbone returns numerical distributions for independent choice, score and noul questions without generating answer text.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
andrueandersoncs/visual-jevPython
andrueandersoncs/visual-jevPythonA Hermes Agent plugin that puts typed Jev judgments in front of the background skill curator: per-pair relation choices and coverage probabilities feed star-shaped merge plans, while core keeps ownership of every mutation.
Jev Agents & MCP#community#catalog-import#madewithjev
Community
anpicasso/hermes-jev-curatorPython
anpicasso/hermes-jev-curatorPythonA self-hosted MCP gateway on Cloudflare Workers that wraps the TypeSafe/Jev System One model into nine judgment tools for intent routing, retrieval reranking and batch semantic work, with a D1-backed admin console.
Jev Agents & MCP#community#catalog-import#logicrw
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.