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 on Hugging Face under the Apache 2.0 license. It also introduced a reinforcement-learning product for customers to fine-tune Clef. The benchmark results and speed comparisons cited in the announcement are Cloudflare’s own evaluations.

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, aiming to make structured, model-generated decisions easier to put into software workflows.

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 decision models as systems that return structured classifications and probabilities from provided inputs, rather than open-ended text. In its example, a customer-support message could be classified by urgency and routed to a team, escalated, or sent to a person for review. The company says Clef can also handle images and has a 64,000-token context window; it contrasts those features with Typesafe AI’s Jev, which Cloudflare says has a 32,000-token window and currently supports text classification.

Cloudflare says Clef is currently leading on the Jev Decision Index, but that result is based on the benchmark presented by the company and should be read in that context. In a separate internal website-classification test, Cloudflare reports that Clef took 2.2 seconds to fetch, render and classify a site using Browser Run. The company compared this with 4.7 seconds for its fastest general model in that workflow, gpt-oss-120b, which it says returned two classifications. These are reported results from Cloudflare’s test, not an independent comparison.

The company also reports that its models performed well across a range of decision-making benchmarks and beat Jev in three of four workflows in Typesafe’s evaluation suite. Cloudflare says Clef-flash was substantially faster in its latency evaluations. It has not provided independent verification in the source material, and results can depend on benchmark design and the specific tasks being tested. The models are accessible through Workers AI, while the Hugging Face release allows users to run and experiment with them outside Cloudflare’s hosted service.

At a glance
announcementWhen: Announced in the Cloudflare Blog; the s…
The developmentCloudflare released its Clef decision models and announced a reinforcement-learning platform for fine-tuning Clef.

Structured Decisions for AI Workflows

Decision models are intended to handle a narrower part of an AI workflow than general-purpose language models: returning bounded, machine-readable outputs that software can act on. That can suit routing, classification and escalation tasks where an application needs a consistent answer in a known format, rather than a generated explanation or broad conversation.

For developers, the combination of a hosted endpoint, an open-source release and a fine-tuning product offers different ways to test the models. The open license permits local experimentation, while the RL product could let customers adapt a model to their own decisions. How well that works in practice will depend on customer data, evaluation methods, latency requirements and safeguards for consequential decisions. Cloudflare’s announcement does not establish that automated decisions can replace human review in every use case.

Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) - Metallic Black

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Bose QuietComfort Wireless Noise Cancelling Headphones, Sandstone

Bose QuietComfort Wireless Noise Cancelling Headphones, Sandstone

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray

As an affiliate, we earn on qualifying purchases.

How Clef Fits Decision Models

Cloudflare presents Clef as part of a growing category of decision models, citing Typesafe AI’s Jev System One as a recent example. The category is distinct from general language models in the company’s framing: rather than generating open-ended responses, a decision model is asked to classify inputs against a task and return outputs that software can consume.

Cloudflare says its Threat Intelligence team has tested Clef to classify website domains, using Browser Run to fetch and render pages before classification. The example illustrates a potential use within an existing workflow, but it is not evidence of performance across all domains or security scenarios. The company also says hosting Clef on Workers AI lets it use GPUs across its infrastructure to reduce network latency; the announcement does not quantify that benefit separately from model latency.

“Clef is currently the leader when evaluated against the Jev Decision Index.”

— Cloudflare

Benchmark Scope and Fine-Tuning Details

The source does not specify the publication date, pricing, availability limits or technical requirements for the new reinforcement-learning product. It also does not detail how customers provide training data, what controls are available during fine-tuning, or how the resulting models will be evaluated and monitored.

Cloudflare’s reported rankings and latency figures reflect its stated tests. The announcement does not describe independent replication, and performance may vary with hardware, prompts, input length and task design. It is also unclear which decision tasks are appropriate for automated handling without human review, or how confidence thresholds and errors should be managed in high-impact settings.

Availability and Customer Evaluations

Developers can access Clef and Clef-flash through Workers AI and find the open-source releases on Hugging Face, according to Cloudflare. The company’s announcement points readers to a live Jev Decision Index demonstration for benchmark results. Customers interested in tailoring Clef can look for further details from Cloudflare about access to the RL product and its fine-tuning process.

The next useful evidence will come from documentation on product availability and from evaluations outside Cloudflare’s own tests. Those should clarify how fine-tuned versions perform on customer-specific tasks, what the trade-offs are between Clef and Clef-flash, and when a system should defer a decision to a person.

Key Questions

What did Cloudflare announce?

Cloudflare released Clef and Clef-flash, decision models hosted on Workers AI and available on Hugging Face under an Apache 2.0 license. It also announced an RL product for fine-tuning Clef.

What is a decision model?

In Cloudflare’s description, it takes inputs and returns structured classifications with probabilities, which an application can use to route work, escalate a case or request human review.

Can Clef process images?

Cloudflare says Clef has a vision encoder and can classify visual content. The announcement does not provide independent results for image-based tasks.

Are Cloudflare’s benchmark results independently verified?

The source presents the results as Cloudflare’s evaluations and does not report independent verification. The reported rankings and timings should be understood as company claims about its tests.

What is not yet known about the fine-tuning platform?

The announcement does not spell out pricing, access requirements, data handling, evaluation procedures or safeguards for the RL product. Cloudflare has not supplied those details in the cited material.

Source: hn

You May Also Like

Running Gemma 4 26B At 5 Tokens/sec On A 13-Year-old Xeon With No GPU

A 13-year-old Xeon server successfully runs Gemma 4 26B at 5 tokens/sec without a GPU, demonstrating impressive efficiency on aging hardware.

Discover 5 AI-Driven Google Search Hacks To Upgrade Your Home Decor

Google introduces five AI-driven search tools to enhance home decor planning, from visualizing furniture to price comparison, with varying availability.

What’s in a GGUF, besides the weights – and what’s still missing?

A detailed analysis of GGUF files for language models, highlighting confirmed features and what still needs to be developed for full functionality.

Can I Opt Out Of My Input Or Output Data Being Used For Training?

Exploring whether individuals can refuse to have their input or output data used to train AI models, amid rising privacy concerns and industry debates.