TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Cloudflare says it has released two decision models, Clef and Clef-flash, through Workers AI and under an Apache 2.0 license on Hugging Face. It also introduced a reinforcement-learning product for customers to fine-tune Clef; independent validation, product details and availability terms were not included in the source report.
Cloudflare says it has released Clef and Clef-flash, two decision models available through its Workers AI service and as open-source weights on Hugging Face under an Apache 2.0 license. The company also introduced a reinforcement-learning product that it says will let customers fine-tune Clef for their own use cases, positioning the models for structured decisions inside software and AI-agent workflows.
Cloudflare describes a decision model as a system that classifies inputs and returns structured answers with probabilities. In its example, a support message could be labeled for urgency and assigned to a team, allowing software to route a ticket, escalate it or send it to a person. The model is intended to produce bounded outputs for a defined decision, rather than the open-ended text and tool use associated with a general-purpose large language model.
The company says Clef can process images as well as text and has a 64,000-token context window. It describes Clef-flash as a faster variant. Both are compatible with the Jev API, according to Cloudflare, which says this allows users to experiment with the hosted models using that interface. The weights are also available for local experimentation under the stated license.
Cloudflare reported an internal website-classification test using its Threat Intelligence team’s workflow and Browser Run. It says Clef fetched, rendered and classified a website in 2.2 seconds, compared with 4.7 seconds for the company’s fastest general model in that workflow, gpt-oss-120b. Cloudflare also says Clef returned more classifications in that comparison. These figures are company-reported results from a particular test, not a general performance guarantee.
Structured Decisions for Agent Workflows
Decision models target a practical gap in agent systems: software often needs a predictable, machine-readable choice—such as a category, route or escalation—rather than a free-form response. If a model can return a decision and confidence information quickly, developers may be able to put it directly into workflows that currently rely on rules, slower model calls or human review.
The potential benefit comes with a responsibility. A probability or classification is not proof that an answer is correct, and automated routing can carry consequences when inputs are ambiguous or a model is wrong. Cloudflare’s description includes deferring to a person as an option, but the announcement does not establish that every use case can safely remove human review. Customers will need to test performance on their own data and decide what decisions require escalation.
The release also gives developers a choice between a hosted service and locally run models. The Apache 2.0 release can support inspection and experimentation, while Workers AI offers a managed route to use the models. Cloudflare’s new fine-tuning product could let customers adapt a model to specialized tasks, although the company has not provided enough detail in the supplied report to evaluate its cost or operational requirements.
open-source decision model software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Clef Fits the Decision-Model Field
Cloudflare’s announcement follows attention around decision models, including Typesafe AI’s Jev System One. The company presents these systems as distinct from general-purpose language models: they are built to return bounded classifications and other structured outputs, while language models can generate open-ended text and handle a broader range of tasks. The categories can overlap in real applications, but the distinction helps explain the intended role for Clef.
Cloudflare compared Clef and Clef-flash with several models across benchmarks covering tool retrieval, API use, classification and other tasks. Its report says Clef led the Jev Decision Index at the time of publication and that the Clef models beat decision-model competitors on latency across 43 evaluations, with an exception for Laya. It also reported that Clef outperformed Jev in three of four evaluations drawn from Typesafe’s suite. These are vendor-reported benchmark results; the supplied material does not describe independent replication or all evaluation conditions.
The company’s domain-classification example illustrates the intended deployment setting: a model receives information gathered by another tool, assigns categories and returns results that downstream code can use. Cloudflare says its models are hosted on Workers AI and can use its edge infrastructure. That is the company’s explanation for potential network-latency benefits, not a measurement that can be assumed for every location or deployment.
“A decision model makes classifications to help agents decide how to act, based on certain probabilities.”
— Cloudflare
As an affiliate, we earn on qualifying purchases.
Benchmark Scope and Product Terms
The announcement does not provide independent evaluations or enough methodological detail to determine how well the reported benchmark results generalize to other datasets and production settings. The 2.2-second versus 4.7-second comparison concerns one Cloudflare-described workflow, and latency can vary with hardware, network conditions, input size and configuration.
Details of the reinforcement-learning product are also limited in the supplied report. It does not specify when the fine-tuning service will be generally available, what it will cost, which training methods or controls it offers, or how customers’ data will be handled. The source also does not state the models’ exact release date. Those points remain unconfirmed here.
structured decision classification tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing and Fine-Tuning Access
Developers can examine the model releases on Hugging Face and try the hosted versions through Workers AI, according to Cloudflare. The next practical test is whether Clef and Clef-flash perform reliably on customer-specific tasks, including cases where classifications are uncertain or errors could affect users.
Further information from Cloudflare about the fine-tuning product, including access, pricing, data handling and evaluation guidance, would clarify how customers can adapt Clef and deploy it responsibly. Until then, the announcement confirms the models’ release and the introduction of the product, but not the full terms or measured outcomes of customer fine-tuning.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are Clef and Clef-flash?
They are decision models released by Cloudflare. The company says they return structured classifications and probabilities for use in software workflows; Clef-flash is described as the faster variant.
Where can developers access the models?
Cloudflare says the models are hosted on Workers AI and their weights are available on Hugging Face under an Apache 2.0 license for local experimentation.
What does the new reinforcement-learning product do?
Cloudflare says the product will let customers fine-tune Clef for their use cases. The supplied announcement does not specify its availability, price, or technical and data-handling terms.
Are the reported speed and benchmark results independently verified?
The supplied material presents them as Cloudflare’s own results. It does not include independent replication or enough methodological detail to establish how broadly the findings apply.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
