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 Clef and Clef-flash, decision models hosted on Workers AI and available under an Apache 2.0 license on Hugging Face. It also introduced a reinforcement-learning product for tailoring Clef to customer use cases. Performance figures in the announcement come from Cloudflare’s own evaluations and have not been independently verified here.
Cloudflare has released Clef and Clef-flash, two decision models hosted on its Workers AI platform, and introduced a reinforcement-learning product that customers can use to fine-tune Clef for their applications. The company says the models are available on Hugging Face under an Apache 2.0 license, allowing developers to run and experiment with them outside Cloudflare’s hosted service.
Cloudflare describes a decision model as a system that returns structured classifications and probabilities for software to use when choosing what to do next. For example, a support workflow could use one to estimate whether a message is urgent and which team should receive it. The output can guide an automated action or send an uncertain case to a person. That differs from general-purpose large language models, which can produce open-ended text and tool calls but are less deterministic, according to the company.
The models are Jev-API compatible, Cloudflare says, so developers familiar with Jev-style decision models can try them through the hosted offering. Cloudflare also says Clef has a vision encoder and a 64,000-token context window; it contrasts those features with Jev’s text-only capability and 32,000-token context window. The new RL product is intended to let customers adapt Clef to their own use cases, though the announcement provides limited detail about how the fine-tuning service works.
Cloudflare reported results from its internal evaluations, including tests against other decision models and workflows in Typesafe’s evaluation suite. It said Clef led the Jev Decision Index evaluation and that its models beat competing decision models on latency across 43 benchmarks, with an exception for Laya. These are company-reported benchmark results, not an independent assessment. In a domain-classification example, Cloudflare said Clef took 2.2 seconds to fetch, render and classify a website using Browser Run; it reported 4.7 seconds for its fastest general model, gpt-oss-120b, on the same workflow.
Where Clef Fits in Agent Workflows
The release targets developers building systems that need to make bounded, repeatable decisions as part of a larger workflow. A model that returns typed categories with confidence estimates can make it easier for code to route a support request, label a domain or decide when a person should review a case. Cloudflare’s pitch is that such decisions can sit in an agent’s operational path without asking a general language model to generate an unrestricted answer each time.
That distinction matters most where latency and consistency affect the usefulness of automation. Cloudflare’s threat-intelligence example concerns classifying websites, including the possibility of flagging suspected phishing. The company says Clef returned multiple classifications more quickly than the general model in its test. If those results hold for customers’ workloads, teams could have a faster way to triage inputs, while still setting their own thresholds for escalation. The announcement does not establish that Clef is suitable for every high-stakes decision or that its confidence scores alone are sufficient grounds for automated action.
Open licensing and the fine-tuning offer address different adoption needs: developers can download and run the models, while customers can use Cloudflare’s hosted service and adapt Clef to particular tasks. The practical value will depend on the quality of customer-specific results, the cost and accessibility of tuning, and how well the models perform beyond the evaluations Cloudflare selected.
As an affiliate, we earn on qualifying purchases.
From Jev to Cloudflare’s Clef
The announcement follows wider interest in decision models, including Typesafe AI’s Jev System One. Cloudflare frames the category as complementary to general-purpose language models rather than a replacement for them: a decision model classifies information into constrained outputs, while an LLM can reason and generate more open-ended responses. In a combined workflow, the choice of model would depend on whether the task calls for a defined decision or broader generation.
Cloudflare says its Threat Intelligence team has been testing Clef to classify website domains, using Browser Run to fetch and render pages before classification. The company also published comparison tables for several benchmark tasks, including API use, customer service, security incidents and classification. The reported scores vary by task: Clef-flash was strongest on some listed evaluations, while Jev scored higher on others. Cloudflare’s overall claims about leaderboard position and latency should be read alongside those task-specific results and the company’s role as the model developer.
““Today, we’re releasing two Cloudflare-trained decision models, Clef and Clef-flash, hosted on Workers AI.””
— Cloudflare
reinforcement learning fine-tuning platform
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Questions on Tuning and Evaluation
The announcement does not specify the availability, pricing or technical limits of the reinforcement-learning product, nor does it explain what customer data or computing resources fine-tuning will require. It is also unclear which Clef versions customers can tune, how the resulting models can be deployed, or whether the fine-tuning service is generally available.
Cloudflare’s benchmark tables and latency comparisons are company-reported. The source does not describe independent verification, provide enough detail here to reproduce every test, or establish how results will transfer to different prompts, datasets, deployment settings or real-world error costs. The domain example is a single reported workflow, not evidence of general performance across all sites. The source also gives no publication date, so the precise timing of the release and product availability cannot be established from the supplied material.
As an affiliate, we earn on qualifying purchases.
Testing Clef in Customer Systems
Developers can evaluate the Apache 2.0 models on Hugging Face or try the Jev-compatible versions hosted on Workers AI, according to Cloudflare. The company also points users to a live decision-index demonstration for its benchmark results. For prospective users, the next practical step is to test the models on representative inputs and compare accuracy, latency and human-escalation needs with their existing workflow.
Cloudflare has not announced a date for additional model releases or supplied a detailed schedule for the RL product. Further information about access, tuning procedures, cost and customer deployment options will be needed to assess how readily the new service can move from demonstration to production use.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are Clef and Clef-flash?
They are Cloudflare-trained decision models designed to return structured classifications and probabilities for use in software workflows. Cloudflare hosts them on Workers AI and says they are available on Hugging Face under Apache 2.0.
What does the new reinforcement-learning product do?
Cloudflare says customers can use it to fine-tune Clef for their use cases. The announcement does not provide detailed information about access, pricing, data requirements or the tuning process.
How does a decision model differ from a general language model?
A decision model is designed to return bounded, structured outputs, such as categories and probabilities. A general language model is more open-ended and can generate text or tool calls; the two can serve different roles in an application.
Are Cloudflare’s benchmark results independently verified?
The source presents them as Cloudflare’s evaluation results. It does not establish independent verification, so the reported comparisons should not be treated as external confirmation.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
