Urgent.News

What's breaking now, across thousands of outlets.

Science

Cloudflare tries to outplay Jev with open-weight Clef models

Sure, it costs more, but it can handle images and video and it's available on Hugging Face if you have the hardware horsepower to run it locally

Cloudflare tries to outplay Jev with open-weight Clef models

Two weeks after the Jev model captured the attention of the AI community, Cloudflare unveiled its own Clef decision models. Promising to outperform Jev in terms of speed, accuracy, and functionality, Clef consists of two versions: Clef and Clef-flash, the latter being a more compact and quicker iteration of the former. Both models are capable of responding to three types of structured questions: yes/no, multiple-choice, and rankings.

However, Clef distinguishes itself from Jev by employing a unique architecture built on specially fine-tuned Qwen3.8-27B and Qwen3.5-9B backbones, which perform a prefill-only pass during inference.

Cloudflare conducted tests to compare Clef's performance against Jev and other open decision models. Using the Jev Decision Index on Hugging Face, Cloudflare found that Clef is slightly slower than other open models but significantly more accurate. Clef-flash, the smaller variant, is equally accurate as most other models while being far quicker.

Cloudflare also noted that Clef surpassed Jev in three out of four areas during their own benchmark tests, with the only discrepancy being agent trace observability. Moreover, Clef possesses the advantage of not being restricted to text classification, as it can also process images and video. Clef can handle up to a 64k context window, surpassing Jev's limit of 32k tokens - including both the state and the longest individual question.

Accessible through Cloudflare's Workers AI platform, Clef enables faster decision-making by leveraging the company's GPU resources at the edge, thereby reducing network latency. For users who prefer not to incur token costs, Clef can be downloaded from Hugging Face under the same Apache-2.0 license as Qwen. However, the training datasets remain undisclosed, and Cloudflare emphasizes that Clef-flash requires a minimum of 41 GB of VRAM on a GPU while Clef necessitates 85 GB of VRAM for optimal performance.

Despite these hardware requirements, Cloudflare assures users that Clef's API remains fully compatible with Jev, facilitating a seamless transition as either a standalone model or through Cloudflare's hosting option.

Written by urgent.news from The Register Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

This story

This is one outlet's version. Read the fullest account.

Read the original at theregister.com →

More in Science

More from Thursday 1 October →