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, on Workers AI and under an Apache 2.0 license on Hugging Face. It also announced a reinforcement-learning product for customer fine-tuning; the performance figures cited so far come from Cloudflare’s own evaluations.
Cloudflare has released Clef and Clef-flash, two decision models hosted on its Workers AI platform and available as open-source models on Hugging Face under an Apache 2.0 license. The company also announced a new reinforcement-learning (RL) product that it says will let customers fine-tune Clef for their own use cases, adding both model access and a customization option for software that needs structured decisions.
Cloudflare describes a decision model as a system that takes inputs and returns structured classifications with probabilities. In a customer-support example, a model could classify a message as urgent and recommend a team, allowing an application to route or escalate the case, or send it to a human. The approach is aimed at bounded decisions inside workflows, rather than the open-ended text generation associated with large language models.
The company says Clef has a vision encoder for classifying images and a 64,000-token context window, compared with the 32,000-token window it attributes to Typesafe AI’s Jev. Cloudflare says its models are compatible with the Jev API, making them usable with software built for that interface. It presents Clef as the stronger model on the Jev Decision Index, but the supplied announcement does not provide independent validation of that ranking.
Cloudflare also described an internal domain-classification test using its Browser Run tool. It says Clef fetched, rendered and classified a website in 2.2 seconds, while its fastest general model in that comparison, gpt-oss-120b, took 4.7 seconds and returned two classifications. Cloudflare reported multiple scores across decision-model benchmarks, with results varying by task; those figures are company-reported evaluations, not a general guarantee of performance in customer deployments.
Structured Decisions for AI Workflows
The release targets a practical problem for teams building AI agents and automated workflows: converting varied inputs into consistent outputs that software can act on. If a model can classify a request, image or website into defined categories, an application may be able to route work or trigger a next step without asking a general-purpose language model to generate and interpret a longer response.
Cloudflare’s open-source release gives developers the option to run the models locally, while Workers AI offers a hosted route. The company’s new RL product could also let customers adapt Clef to particular tasks, though the announcement does not yet detail how that service works or what resources it requires. The combination matters to developers weighing customization, deployment location, speed and control over model behavior.
Cloudflare’s latency and benchmark results are relevant evidence about its own tests, but buyers will need to compare performance on their own data and workloads. Results can depend on the task, model configuration and surrounding application; the reported website test alone does not establish that Clef will be faster or more accurate in every setting.
As an affiliate, we earn on qualifying purchases.
How Clef Fits Decision Models
Cloudflare’s announcement places Clef in a developing category it calls decision models. The company contrasts these systems with general-purpose language models: decision models are meant to return bounded, typed outputs for a particular workflow, while language models can generate open-ended text and tool calls. In practice, a system may use either kind of model—or combine them—depending on whether it needs a defined classification or broader reasoning.
The announcement discusses Typesafe AI’s Jev as an existing example and says Clef is designed to work through the Jev API. Cloudflare compares model results on a range of benchmarks and says Clef performed well across several of them. Its posted table also shows that no single model leads every listed task, reinforcing that the evaluation depends on the benchmark and metric rather than one overall score.
Cloudflare says the models are hosted on Workers AI, where they can use the company’s distributed infrastructure, and are also being released through Hugging Face. The announcement’s benchmark data and deployment claims are from Cloudflare; the source material does not include an external evaluation or customer case study.
““Today, we’re releasing two Cloudflare-trained decision models, Clef and Clef-flash, hosted on Workers AI.””
— Cloudflare
open-source machine learning models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Fine-Tuning and Evaluation Details
The announcement does not specify when the RL product will be available, how customers will submit training data, what access or pricing terms apply, or which model parameters and outputs can be customized. It also does not describe safeguards, evaluation procedures or how a fine-tuned model can be checked for reliability before use in consequential workflows.
Cloudflare’s comparisons are based on its own benchmark runs, including tests against models named in its post. The supplied material does not include independent replication, detailed test conditions for every result, or evidence that the scores predict outcomes across different customer systems. The company’s 2.2-second website-classification example is a single reported workflow; broader performance remains to be established.
It is also not clear from the announcement whether every aspect of the RL fine-tuning platform will be open-source, or whether it will be a hosted service. The open-source licensing statement applies to the Clef models, not necessarily to the separate fine-tuning product.
As an affiliate, we earn on qualifying purchases.
Availability and Customer Testing
Developers can evaluate Clef and Clef-flash through Workers AI or obtain the models from Hugging Face under the stated Apache 2.0 license. The next practical test is how the models perform on real workflow data, particularly where classification errors can trigger actions or affect customers.
Further details from Cloudflare are needed to establish the RL product’s launch timing, operating requirements and customization process. Customers considering adoption will also need to assess task-specific accuracy, latency, data handling and the option to send uncertain cases to human review. Until those details and independent evaluations are available, the release establishes access to the models and a stated fine-tuning plan, but not their performance across production use cases.
structured decision classification tools
As an affiliate, we earn on qualifying purchases.
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 says both are hosted on Workers AI.
Can developers run the models outside Workers AI?
Cloudflare says Clef and Clef-flash are available on Hugging Face under an Apache 2.0 license, allowing developers to access and run the models outside the hosted service, subject to the license terms.
What does Cloudflare’s RL product do?
Cloudflare says the new reinforcement-learning product will let customers fine-tune Clef for their use cases. The announcement does not specify its release date, pricing, data requirements or technical process.
Are the benchmark results independently verified?
The results described in the announcement are Cloudflare-reported evaluations. The supplied source material does not include independent replication or external verification.
How does a decision model differ from a general language model?
In Cloudflare’s description, a decision model returns bounded, structured outputs, such as a category and probability, while a general language model can produce open-ended text and tool calls. Which approach fits best depends on the task.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
