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 two decision models, Clef and Clef-flash, hosted on Workers AI and released under the Apache 2.0 license on Hugging Face. The company also introduced a reinforcement-learning product for customers to fine-tune Clef; performance figures and benchmark comparisons cited so far come from Cloudflare’s own evaluations.
Cloudflare has released Clef and Clef-flash, two open-source decision models hosted on its Workers AI platform, and announced a new reinforcement-learning product for customers to fine-tune Clef for their own use cases. The models are designed to return structured classifications and probabilities that software agents can use to route work or choose actions, rather than generate open-ended text.
Cloudflare says both models are available through Workers AI and on Hugging Face under the Apache 2.0 license, allowing developers to use the hosted versions or run and experiment with the models themselves. The company describes them as compatible with the Jev API, making it easier for users of that interface to test Clef. Cloudflare says Clef currently leads on the Jev Decision Index, but that result and the related comparisons are company-reported benchmark findings, not independent verification.
In a Cloudflare Threat Intelligence test, the company used Clef with Browser Run to fetch and render websites, then classify their domains. Cloudflare reports the workflow took 2.2 seconds, compared with 4.7 seconds for its fastest general-purpose model in the same test, gpt-oss-120b. It also says Clef returned more classifications. The example included likelihood estimates for categories such as fashion, e-commerce and phishing; these are model outputs, not confirmation that any particular site belongs to a category.
Cloudflare says Clef includes a vision encoder for image classification and a 64,000-token context window, compared with 32,000 tokens for Jev. Its lighter Clef-flash version is presented as a faster option. The company reports that its models beat decision-model competitors on latency across 43 evaluations, with an exception for Laya, which it says is very fast but trades off quality on the reported benchmarks.
Structured Outputs for Agent Decisions
Decision models are intended for a narrower task than general-purpose language models: they map supplied context to typed choices and probabilities that software can act on. For businesses building automated support, security or operations workflows, that can make it easier to route a request, flag a potential risk or send an uncertain case to a person. The model’s output can be used as one part of a workflow, rather than requiring an agent to interpret a long free-form response.
Cloudflare’s release also gives developers a choice between a managed service and open-licensed model files. The hosted option runs on Workers AI, while the Apache 2.0 release gives teams the ability to experiment locally. The announced fine-tuning product could let customers adapt the model to their own decision tasks, though the supplied announcement does not describe its practical limits or operating requirements.
The reported speed comparison may matter where a model sits in a system’s real-time “hot path,” but the result is specific to Cloudflare’s test setup. It does not establish that Clef will be faster or more accurate across other hardware, prompts, datasets or production workloads.
As an affiliate, we earn on qualifying purchases.
How Clef Differs From General Models
Cloudflare frames decision models as a category distinct from large language models. In its description, a decision model returns constrained, structured outputs for a specified task, while a general LLM can produce open-ended text and tool calls. A support message, for example, could be classified by urgency and destination team, leaving an application to route it or request human review.
The announcement follows recent attention to Typesafe AI’s Jev System One and other decision-model products. Cloudflare says it evaluated Clef against Jev and additional systems using benchmark tasks for areas including tool retrieval, API use, customer service and security incidents. Its published tables show Clef and Clef-flash leading on some measures and trailing on others, underscoring that results vary by task and metric. The company also says Clef-flash performed particularly well on speed, while scoring differently from Clef across the listed evaluations.
““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.
Benchmark Scope and Product Details
The announcement does not provide an independent evaluation of the benchmark results, and the reported rankings should be read as Cloudflare’s own measurements. The supplied material does not specify all evaluation conditions, such as hardware, repeat counts or how each benchmark maps to production performance. It also does not establish whether the Threat Intelligence timing will generalize to other sites or workflows.
Details about the new RL fine-tuning product remain limited. Cloudflare does not state in the supplied announcement when it will be available, how customers will access it, what data or compute it requires, or what controls will govern fine-tuning. The announcement also does not quantify the models’ error rates in real-world deployments or explain how customers should set thresholds for deferring uncertain decisions to people.
structured decision-making AI tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Access, Fine-Tuning and Evaluation
Developers can test the hosted models through Workers AI and access the open releases on Hugging Face, according to Cloudflare. The company points readers to a live Jev Decision Index demonstration for its benchmark results. It has not given a timeline or further implementation details for the reinforcement-learning product in the supplied material.
For customers assessing Clef, the next practical steps are to test it against their own decision tasks, compare both model versions on relevant measures, and decide when an uncertain output should be reviewed by a human. Further information from Cloudflare about the fine-tuning service and independent evaluations would help clarify how the models perform beyond the company’s reported tests.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are Clef and Clef-flash?
They are Cloudflare-trained decision models that return structured classifications and probabilities for use in software workflows. Clef-flash is presented as a faster variant.
Where can developers access the models?
Cloudflare says they are hosted on Workers AI and released on Hugging Face under the Apache 2.0 license for local use and experimentation.
What does Cloudflare’s RL platform do?
Cloudflare says the product will let customers fine-tune Clef for their use cases. The announcement does not give details on availability, requirements or how the service works.
Are the benchmark results independently verified?
The supplied source presents them as Cloudflare’s evaluations. It does not cite an independent audit, so the reported comparisons should not be treated as independently verified.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
