Cloudflare's open weights for steering agents: the Clef family debuts

On October 1, 2026, network services company Cloudflare announced on its official blog, in a post signed by Michelle Chen, the release of Clef and Clef-flash. They are described as "decision" models, with weights distributed on Hugging Face under the Apache 2.0 license and hosted on the Workers AI platform. Unlike traditional generative systems, these tools do not produce free text: their concrete job is to classify data, route requests and assign scores, returned in typed formats along with a confidence index. As Michelle Chen explains in the launch post, "a decision model performs classifications to help agents decide how to act, based on certain probabilities". Both models are built on the Qwen architecture (specifically Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash, post-trained with rank-256 low-rank adapters), have a 64,000-token context window and include a vision encoder that lets them process images too.

In this niche, where TypeSafe's Jev System One was already present and Amazon's Strands Decider 2B arrived the same day, Cloudflare is betting on latency: according to the company, its models beat the other decision models, "except Laya, which is very fast but sacrifices quality". The stated figures are a median latency of 209.3 milliseconds for Clef, with a 95th percentile of 238.6 milliseconds, and 38.8 milliseconds for Clef-flash, with p95 at 122.4 milliseconds; for Jev System One the stated median is 524.1 milliseconds. Clef is also compatible with the API of Jev System One, the first decision model to reach the market. On usage costs on Workers AI, the specialist outlet Developers Digest reports rates of $0.24 per million input tokens for Clef and $0.09 for Clef-flash, with no charge for output tokens. The company also offers a customization service based on reinforcement learning, run by its own engineering team, noting that "when you fine-tune a model, you may give up some general performance in exchange for greater accuracy in a specific domain", though no release date has yet been given for a self-service interface.

The transparency of the operation, however, has some methodological grey areas. The benchmarks Cloudflare published, picked from the Jev Decision Index evaluations — such as Clef's 98.47% and Clef-flash's 98.76% on BFCL case exact — are measured by the company itself, with no independent verification. Nor does the post announce the release of the training data or pipeline; it merely mentions "internal synthetic datasets": an absence that places this release among open-weight distributions rather than fully reproducible open source projects. The official post also lacks direct comparisons on general reasoning, such as GPQA Diamond or MMLU-Pro, areas in which, according to Developers Digest, Jev System One is clearly ahead.

Replacing text generation with a stream of typed probabilities is an excellent engineering move to cut latency in commercial agents, and the company's stated guarantee that "we do not read, store or train models on your requests or responses" answers a real enterprise security need. Still, an automated decision-maker's effectiveness is measured by the complexity of the forks it can handle, and on that front the published numbers don't tell us anything yet.

— Olya