Cloudflare’s Latest Challenger: The Clef Models
Two weeks following the explosive debut of the Jev model, Cloudflare has unveiled its own duo of “Clef” decision models, claiming they are more intelligent and quicker than Jev, while also being open weight and operable locally. The Clef model lineup was introduced by Cloudflare on Thursday, featuring two variants: Clef and Clef-flash, the latter being a slightly smaller and quicker iteration of the original model.
Grasping the Fundamentals
Essentially, they operate similarly to TypeSafe’s Jev, capable of addressing three categories of bounded, structured inquiries: Yes/no, multiple choice, and rankings. However, this is where substantial distinctions arise, as Clef is not only constructed differently but, if Cloudflare’s benchmark assertions prove accurate, also seems to outperform Jev and various other decision models across multiple tests.
The LLM Infrastructure
To begin with, Clef is built on an LLM infrastructure. Cloudflare states that Clef utilizes specially post-trained, frozen versions of Qwen3.8-27B and Qwen3.5-9B for Clef and Clef-flash, respectively, with the Qwen infrastructure implementing a prefill-only pass during inference. Clef remains speedy – indeed, faster than Jev – and processes choices in parallel after the prefill-only pass. Jev’s underlying architecture remains undisclosed, as TypeSafe has kept that information confidential.
Performance and Features
Velocity and Precision
Regarding its speed and functionality, Clef operates quickly. Cloudflare tested it against Jev and several other open decision models using the Jev Decision Index found on Hugging Face, and the company’s own assessment indicates that Clef is marginally slower than other open models but exhibits greater accuracy, with Clef-flash achieving similar accuracy to most others, yet at a significantly faster rate. In fairness to competitors, Cloudflare self-reported its own scores against the benchmark, which have not yet been replicated for official ranking on the Decision Index.
Expanding Beyond Text
Even if it was somewhat slower or less accurate, Clef has a major advantage over Jev: It’s not confined to text classification – it is also capable of processing images and videos. Moreover, Clef accommodates a 64k context window. Jev can handle up to 64k tokens within a request, although its state plus the longest individual question is restricted to 32k.
Accessibility and Pricing
Clef is accessible directly from Cloudflare, hosted on Workers AI, which the company claims enhances the models’ speed because “we’re able to utilize our GPUs at the edge, resulting in low network latency and quicker decisions.” For users who prefer to avoid the token fee (Clef costs $0.24 for every million tokens – almost six times Jev’s rate at $0.042/M), Clef can also be downloaded from Hugging Face, and is open weight under the same Apache-2.0 conditions as Qwen.
Executing the Models
While referred to as “open source” in the announcement, Cloudflare AI Platform group product manager Michelle Chen confirmed to GadgetLad that its training datasets are not publicly available. Concerning hardware compatibility, Chen informed us that Clef-flash will operate on any GPU with a minimum of 41 GB of VRAM, while Clef necessitates 85 GB of VRAM on a GPU to function. “This is based on single concurrency and a 64k context window,” Chen specified. There’s no need to overhaul your decision model architecture for Clef either – its API is fully compatible with Jev, permitting it to serve as a drop-in substitute if you want to give it a try, either locally or through Cloudflare’s hosting service.
Conclusion: The Clef Ensemble Has Arrived
Cloudflare’s Clef models have made a striking entry, poised to eclipse the Jev models with their impressive features and remarkable speed. Whether you aim to elevate your text, image, or video classification capabilities, Clef could very well be the exciting new duo of models you require. It’s time to give these dynamic models a whirl and discover if they can truly hit the right notes!
