AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Cloudflare announced two decision models, Clef and Clef-flash, hosted on Workers AI and released as open-source models under the Apache 2.0 license. The company also introduced a reinforcement-learning product for fine-tuning Clef, but has not provided details on its availability, pricing or technical operation in the announcement.

Cloudflare has released Clef and Clef-flash, two decision models hosted on its Workers AI platform, and says both are available on Hugging Face under an Apache 2.0 license. The announcement also introduces a reinforcement-learning product for customers to fine-tune Clef, aimed at adapting the models to specific tasks such as classification and workflow routing.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Cloudflare describes a decision model as a system that returns structured classifications and probabilities so software can route work, trigger an action or refer a case to a person. A support message, for example, could be classified by urgency and assigned to a team. The company contrasts this narrower, typed-output role with large language models, which can generate open-ended text and tool calls but may be less predictable for fixed decisions.

The models are available through Workers AI and can also be run locally from their Hugging Face releases, according to Cloudflare. The company says Clef accepts images as well as text and has a 64,000-token context window; it contrasts those features with Typesafe AI’s Jev, which the post says is text-only and has a 32,000-token window.

Cloudflare’s post reports results from its own evaluations, including tests drawn from the Jev Decision Index and Typesafe’s evaluation suite. It says Clef led on the Jev Decision Index and that the Clef models beat other decision models on latency across 43 evaluated benchmarks, with an exception for Laya. These are company-reported benchmark results, not an independent assessment. The post also describes an internal domain-classification test: Clef took 2.2 seconds to fetch, render and classify a website, compared with 4.7 seconds for gpt-oss-120b in the same workflow. Cloudflare says its model returned more classifications in that comparison.

At a glance
announcementWhen: Announced in a Cloudflare Blog post; pu…
The developmentCloudflare released Clef and Clef-flash decision models and announced a new reinforcement-learning product to fine-tune Clef for customer use cases.

Where Structured Decisions Fit

Decision models target a specific part of an AI workflow: turning collected information into bounded outputs that software can act on. If a model can reliably classify a request, domain or incident, an application may automate routine routing while reserving uncertain or consequential cases for human review. That could reduce delays in high-volume processes, though the practical value depends on accuracy, the cost of errors and how each system handles uncertain predictions.

Cloudflare’s release also gives developers an option beyond using a hosted API alone. The Apache 2.0 release allows users to inspect, run and adapt the models under that license, while Workers AI provides a hosted route. The proposed fine-tuning product points toward adapting a general decision model to an organization’s categories and procedures. The announcement, however, does not establish how well that approach works for customers or what safeguards and operational requirements it will include.

Products Worth Considering

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

As an affiliate, we earn on qualifying purchases.

From Jev Buzz to Clef

Cloudflare frames the release as a response to recent attention around decision models, including Typesafe AI’s Jev System One. Its post says conventional classifiers often require retraining when categories change, while decision models can handle a range of inputs without repeatedly adding classification categories through retraining. That is Cloudflare’s description of the approach, not a guarantee that models will generalize to every new task without changes or evaluation.

The post positions Clef alongside, rather than as a replacement for, general-purpose language models. A decision model produces constrained outputs for a known decision point; an LLM can generate broader responses and use tools. Cloudflare says its own Threat Intelligence team has tested Clef to categorize websites, including estimates for categories such as fashion, ecommerce and phishing. The example illustrates the intended workflow, but the post does not give a full account of the test set or error rates for that deployment.

““Today, we’re releasing two Cloudflare-trained decision models, Clef and Clef-flash, hosted on Workers AI.””

— Cloudflare

Fine-Tuning Details Remain Sparse

The announcement does not specify the RL product’s release date, pricing, access conditions or technical workflow. It also does not explain what customer data or computing resources fine-tuning will require, how resulting models can be deployed, or what evaluation and safety controls will be offered.

Benchmark figures and the domain-classification comparison come from Cloudflare’s own testing. The supplied report does not include independent validation, complete methodology for every benchmark, or enough deployment detail to establish how Clef will perform across different customer environments. The company’s claim that Clef leads on the Jev Decision Index should be read as a result on that stated benchmark, not as a broad ranking across all decision-model tasks.

Availability and Evaluation Ahead

Developers can try the models through Workers AI or obtain the open-source releases from Hugging Face, according to Cloudflare. The company says the models are Jev-API compatible, which may let developers experiment through an existing interface, though the post does not provide further implementation or migration details in the supplied material.

For customers interested in adapting Clef, the next concrete information to watch for is documentation on the reinforcement-learning product, including access, pricing, training requirements and deployment options. Further independent testing and customer deployments will also help clarify whether the reported speed and benchmark results carry over to real workflows, and how teams should handle low-confidence or high-impact decisions.

Key Questions

What are Clef and Clef-flash?

They are Cloudflare-trained decision models designed to return structured classifications and probabilities. Cloudflare hosts them on Workers AI and says they are also available on Hugging Face under the Apache 2.0 license.

What does a decision model do?

It classifies information into defined outputs that software can use to route a task, trigger an action or request human review. Cloudflare gives customer-support triage and website categorization as examples.

How does Clef differ from Jev?

Cloudflare says Clef includes a vision encoder and a 64,000-token context window, compared with Jev’s text-only capability and 32,000-token window. The company also reports comparative benchmark results, which have not been independently verified in the supplied material.

What is Cloudflare’s new RL product?

It is a reinforcement-learning product intended to let customers fine-tune Clef for their own use cases. The announcement does not explain its availability, pricing, training process or deployment terms.

Source: hn

You May Also Like

A Closer Look At Scaling Storage For Over 1 Billion ChatGPT Users

OpenAI’s headline describes scaling online storage for more than 1 billion ChatGPT users, but the available details do not verify the figure or explain the work.

AI Automation Software For Small Business: What To Know Before You Choose

A small-business guide recommends testing one low-risk AI workflow for two weeks before deciding whether to keep an automation tool.

AI Narration, In-House Models: What Went Into GewerkTon’s New Film

“Where GewerkTon works” is a three-minute film through 64 sample buildings – homes, renovations, schools, hospitals, industry, energy, transport and data centres. Every building, plan and model was made by GewerkTon; all projects are fictional.

Opus 5.5 Agents Discover Two Room-temperature Magnetic Semiconductor Candidates

Vals AI reports two computationally predicted candidates for room-temperature magnetic semiconductors, but experimental confirmation is still needed.