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 announced Clef and Clef-flash, open-source decision models hosted on Workers AI and released on Hugging Face under Apache 2.0. It also introduced a reinforcement-learning product for customers to fine-tune Clef; performance figures in the announcement are Cloudflare’s own benchmark results.
Cloudflare has released Clef and Clef-flash, two decision models available through its Workers AI service and as open-source models on Hugging Face under an Apache 2.0 license. The company also announced a new reinforcement-learning product that lets customers fine-tune Clef for their own use cases, targeting developers who want AI systems to return structured decisions for software workflows.
Cloudflare describes decision models as systems that classify inputs and return structured answers with probabilities, rather than producing open-ended text. In its example, a model could assess a customer-support message for urgency and recommend a team, allowing code to route or escalate the request—or send it to a person. Cloudflare says Clef can process text and images, and has a 64,000-token context window; it contrasts those features with Typesafe AI’s Jev model, which Cloudflare says supports text classification and a 32,000-token window.
Cloudflare reports that it tested Clef on website classification for its Threat Intelligence team. In one workflow using Browser Run to fetch and render a site, the company says Clef took 2.2 seconds to classify it, compared with 4.7 seconds for its fastest general model in that test, gpt-oss-120b. Cloudflare says Clef returned more classifications in the comparison. The figures describe the company’s stated test and do not establish that the same speed or output will apply to other workloads.
The announcement also presents benchmark results comparing Clef and Clef-flash with other decision models. Cloudflare says its models performed competitively across evaluations, and that Clef-flash was faster in its latency tests. The company’s table includes mixed results by task: for example, Clef leads some listed measures, while Jev, Diffusion or Clef-flash score higher on others. Cloudflare says customers can use its new RL product to fine-tune Clef, but the post does not detail its availability, pricing, or training process.
Structured Decisions for AI Workflows
Decision models are aimed at a different job from general-purpose language models: returning bounded, typed outputs that software can act on. If such outputs are reliable for a particular task, teams may be able to route requests, classify domains or select actions without asking a language model to generate a longer response. The model’s probabilities can also support workflows that defer uncertain cases to a human.
That approach could matter in settings where response time and predictable formats affect downstream systems, including support triage and security review. Cloudflare’s example points to a possible latency advantage in one website-classification workflow, but it is a company-reported test, not evidence of a general performance gain. The open license gives developers the option to inspect, run and adapt the models themselves, while the fine-tuning product signals Cloudflare’s interest in tailoring them to customer-specific decisions.
As an affiliate, we earn on qualifying purchases.
From Jev to Cloudflare’s Clef
Cloudflare’s post frames Clef as part of a growing category of decision models, citing Typesafe AI’s Jev System One as a recent example. The company distinguishes these models from large language models: LLMs can generate open-ended text and tool calls, while decision models are designed to return constrained classifications for a defined workflow. Cloudflare says Clef is compatible with the Jev API, which may make it easier for users to experiment with the hosted models.
The company’s comparisons draw on several evaluation sets, including the Jev Decision Index and Typesafe’s evaluation suite. Cloudflare reports that Clef models beat decision-model competitors on latency across 43 benchmarks, with an exception for Laya, which it describes as faster but lower-scoring on quality measures. Those results are presented by the model provider; the post does not describe an independent evaluation. Clef and Clef-flash are available through Workers AI, while their Apache 2.0 release supports local experimentation.
““Today, we’re releasing two Cloudflare-trained decision models, Clef and Clef-flash, hosted on Workers AI.””
— Cloudflare
As an affiliate, we earn on qualifying purchases.
Availability and Results Still Need Detail
The announcement does not specify when the reinforcement-learning product will be generally available, how customers will access it, what it will cost, or which customization controls it will offer. It also does not explain the training data, safeguards, or evaluation process for customer fine-tuning. Those details will affect how teams judge the product’s fit for sensitive or high-impact decisions.
The performance claims should also be read as Cloudflare-reported results. The post lists benchmark scores and describes selected tests, but does not establish independent replication or provide enough detail here to assess every test condition. It is not yet clear how Clef will perform across customer datasets, languages, or workflows beyond the examples and evaluations Cloudflare cites.
reinforcement learning platform for AI
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Product Access and Independent Testing
Developers can try the hosted models through Workers AI or access the releases on Hugging Face under Apache 2.0, according to Cloudflare. The next useful milestones are fuller documentation for the RL fine-tuning product, details on its availability and terms, and evidence from customers or independent evaluators on how the models perform in real workflows. Until then, teams considering Clef will need to test it against their own tasks and decide when its outputs should trigger action or human review.
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 classify inputs and return structured outputs. Cloudflare says Clef can process images as well as text, and both models are hosted on Workers AI.
Where can developers get the models?
Cloudflare says Clef and Clef-flash are available through Workers AI and on Hugging Face under an Apache 2.0 license for local use and experimentation.
What does the new RL product do?
Cloudflare says its reinforcement-learning product will let customers fine-tune Clef for their use cases. The announcement does not give details on launch timing, pricing, or the fine-tuning workflow.
Are the reported speed results independently verified?
The cited figures come from Cloudflare’s own announcement and tests. The post does not report independent verification, so results for other models, tasks, or customer environments remain to be established.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
