Cloudflare Releases Clef Decision Models to Rival TypeSafe's Jev

Cloudflare has introduced Clef and Clef-flash, two open-weight decision models designed to compete directly with TypeSafe’s Jev model, which launched two weeks ago, according to a report by The Register (https://www.theregister.com/ai-and-ml/2026/10/01/cloudflare-tries-to-outplay-jev-with-open-weight-clef-models/5300649). Like Jev, both Clef models are built to process bounded, structured queries across three formats: binary yes-or-no choices, multiple-choice questions, and item rankings.

Unlike TypeSafe, which has kept Jev's underlying architecture private, Cloudflare built Clef on top of open-source large language model backbones. The standard Clef model uses a post-trained, frozen version of Qwen3.8-27B, while Clef-flash runs on Qwen3.5-9B. During inference, the Qwen foundation model runs a prefill-only pass, after which choices are scored in parallel.

A key operational difference between the architectures is input modality. While Jev is limited to text classification, Clef processes text, static images, and video. Clef also supports a 64,000-token context window. While Jev can handle up to 64,000 tokens across an entire request, its state plus its longest individual question is capped at 32,000 tokens.

Performance figures for the models currently rest on vendor-run benchmarks. Cloudflare tested Clef against the Jev Decision Index on Hugging Face, claiming that standard Clef is slightly slower but more accurate than competing open decision models, while Clef-flash matches accuracy at higher speed. When tested against TypeSafe's benchmarks, Cloudflare reported that Clef beat Jev in three out of four evaluation categories, falling short only on agent trace observability. These scores have not yet been independently verified or added to the official Decision Index rankings.

The models are available directly through Cloudflare's Workers AI platform, using edge GPUs to lower network latency. Cloudflare has also published model weights to Hugging Face under the Apache-2.0 license. However, Cloudflare AI Platform group product manager Michelle Chen told The Register that the models' training datasets have not been made public.

Running the models on private hardware requires significant compute resources. Chen stated that running Clef-flash locally requires at least 41 GB of VRAM, while the larger Clef model demands 85 GB of VRAM, assuming single concurrency and a full 64,000-token context window.

To ease migration, Cloudflare made the Clef API fully compatible with Jev as a drop-in replacement. However, using Cloudflare's hosted Workers AI deployment costs $0.24 per million tokens—nearly six times the $0.042 per million token rate for TypeSafe's Jev.