Cloudflare releases open‑source decision models Clef and Clef‑flash: structured judgments for low‑latency AI agents

News Overview

On October 1, 2026, Cloudflare officially announced, via its corporate blog and the Workers AI marketplace, the availability of two open‑source decision models named Clef and Clef‑flash. The two models are each based on Alibaba‑developed base models, Qwen3.8‑27B and Qwen3.5‑9B respectively. Model weights are released under an Apache 2.0 license, allowing users to download, deploy, and even re‑train the models as needed. This release marks a significant shift, indicating Cloudflare’s move from a traditional CDN provider toward a platform that offers AI agents. By making these models publicly accessible, Cloudflare aims to foster broader adoption of decision‑oriented AI solutions at the network edge.

Technical Background and Model Concept

A “decision model” refers to a type of model that does not generate free‑form text, but instead, for predefined true/false, single‑choice, or rating questions, returns the probability distribution for each possible answer. Compared with large generative models, decision models produce outputs that can be more directly embedded into automation pipelines, thereby reducing post‑processing overhead. The Clef series leverages the instruction‑fine‑tuning technique developed for the Qwen family, preserving the base model’s semantic understanding while adding a probabilistic classification head at the final layer. This design enables developers to obtain immediately usable probability vectors without the need for additional text generation steps.

Key Specifications of Clef and Clef‑flash

• Clef: based on Qwen3.8‑27B, roughly 2.7 billion parameters, supports a 64 k token context window, includes a built‑in visual encoder capable of handling mixed image‑text inputs.
• Clef‑flash: based on Qwen3.5‑9B, roughly 900 million parameters, also supports a 64 k token context window, focuses on pure‑text decision making and delivers faster response times.
• License: Apache 2.0, permitting commercial use and redistribution.
• Deployment: callable directly via the Workers AI API and compatible with the TypeSafe Jev SDK.
• Output format: probability vector or binary score, suitable for ticket triage, domain classification, agent guardrails, and routing decision scenarios.

Performance Tests and Latency

In a series of 43 internal evaluations conducted by Cloudflare, the median latency recorded for Clef was 209.3 ms, while Clef‑flash measured 38.8 ms. By contrast, the TypeSafe Jev model run on the same platform exhibited a latency of 524.1 ms, highlighting a clear advantage for the newly released models. Across different quality benchmarks such as accuracy and recall, each model showed distinct strengths: Clef performed better on high‑complexity image‑judgment tests, whereas Clef‑flash excelled in rapid pure‑text decision tasks. Cloudflare also notes that these figures have yet to be independently verified by third‑party parties. The measurements were taken under controlled load conditions and reflect performance observed at Cloudflare’s edge locations.

Reinforcement Learning Fine‑Tuning Platform

Alongside the model releases, Cloudflare introduced a suite of reinforcement‑learning (RL) fine‑tuning services. The platform was initially built by Cloudflare’s engineering team to perform a baseline fine‑tuning, and it now offers a self‑service interface that lets enterprises apply a second round of fine‑tuning using their own data sets. The fine‑tuning workflow includes designing environment feedback, defining a reward function, and iterating the policy, ultimately producing a decision model that aligns with specific business requirements. The platform continues to run within the Workers AI execution environment, ensuring compatibility with the original base models.

Usage Limits and Risks

Cloudflare explicitly warns that organizations planning to adopt Clef or Clef‑flash must first calibrate probability thresholds on their own data, rather than relying solely on the vendor‑provided benchmarks for high‑risk decisions. The models do not guarantee zero error across all domains; judgments involving legal, medical, or financial compliance still require human review. Although the open‑source license permits modification, failure to follow security best practices could introduce model drift or expose the system to adversarial attacks.

Industry Impact and Future Outlook

For companies operating in Taiwan, the Clef series offers a practical way to deploy low‑latency AI judgments within edge‑computing environments. Customer‑service centers can use the models to route tickets in real time, e‑commerce platforms can instantly classify malicious traffic at the CDN edge, and financial institutions can quickly score transaction risk at the API gateway layer. Should independent third‑party validation confirm the reported performance, it is expected that a growing number of SaaS providers will embed such decision models into their service contracts, further accelerating the adoption of AI agents in localized applications.