Cloudflare Launches Clef Decision Models and Reinforcement Learning Fine-Tuning Platform on Workers AI

Cloudflare has introduced Clef and Clef-flash, two new open-source decision models hosted on its Workers AI platform, along with a reinforcement learning service that allows developers to fine-tune decision models using their own data.

The launch follows weeks of industry interest in so-called decision models, a category popularized by Typesafe AI's Jev System One. Unlike classifier models, which have existed for years, decision models produce bounded, structured outputs quickly and cheaply, and can be slotted into a workflow whenever a system needs to make a call. They can operate across a wide range of inputs without needing constant retraining to add new categories. That sets them apart from large language models, which are open-ended and capable of reasoning and generating text, but are largely non-deterministic.

A decision model essentially helps an automated agent decide how to act. Feed it a customer support message, for instance, and it can return typed answers — is the ticket urgent, which team should handle it — along with probabilities, letting code route the request, trigger an escalation, or hand it off to a human only when necessary.

Cloudflare has already been testing Clef internally on its Threat Intelligence team, using it to classify website domains. Given a domain to analyze through Browser Rendering, Clef can quickly assign probability scores across multiple categories — for example, 95% likely a fashion site, 85% ecommerce, under 1% phishing. The classification took 2.2 seconds, compared with 4.7 seconds for Cloudflare's fastest general-purpose LLM, gpt-oss-120b, which also returned only two classifications in the same test.

The name Clef is drawn from music notation, where a clef symbol establishes the pitch context for the notes that follow — a metaphor Cloudflare says fits a model that defines the context for the actions that come next, with the added bonus of echoing the company's own initials.

Cloudflare says Clef differs from other decision models in several ways. It includes a vision encoder, letting it classify images as well as text, whereas Jev currently handles text only. It also has a larger context window — 64,000 tokens compared with Jev's 32,000 — allowing more input to be considered in a single classification.

On benchmark testing tied to the Jev Decision Index, Cloudflare says Clef currently leads the field, with full results published on a live benchmark demo site. Across a broader set of internal evaluations, Clef and Clef-flash also outperformed Jev in three of four workflow categories, including customer service and security incident classification. On latency, Cloudflare says its models beat most competitors, with Clef-flash posting a median response time of under 40 milliseconds — notably faster than Jev's 524 milliseconds — though one rival model, Laya, was faster still while trading off accuracy.

Because the models run on Cloudflare's own edge infrastructure, the company says response times benefit from reduced network latency, making it practical to place Clef in the real-time decision path for agents working alongside large language models.

Clef produces the same strictly typed outputs as Jev and is built to be API-compatible, so developers can switch over with minimal changes. Cloudflare describes Clef as the more powerful, precision-oriented option and Clef-flash as suited to latency-sensitive decisions. Both models are available through Workers AI, and Cloudflare is also releasing the weights on Hugging Face under an Apache 2.0 license for anyone who wants to run them independently.

On the technical side, Cloudflare says Clef builds on earlier experiments adapting the DiffusionGemma model to produce deterministic probability outputs, work that drew on independent research into the vLLM inference engine. Clef instead uses Qwen as its base model, post-trained for decision-making tasks. During inference, it runs a prefill-only pass and scores valid schema choices in parallel rather than generating text token by token, which Cloudflare says makes it significantly faster than standard autoregressive language models. Training combined label-smoothed cross-entropy with a calibration loss, internal synthetic datasets, and a technique Cloudflare calls Reinforcement Learning for Calibrated Decisions, aimed at improving both accuracy and generalization.

Alongside the model release, Cloudflare is launching a reinforcement learning fine-tuning service. Customers will initially work hands-on with Cloudflare's forward-deployed engineering team, with a self-serve fine-tuning platform planned for later, allowing users to capture data, retrain Clef, and redeploy it on Cloudflare's infrastructure. The company says internal teams are already exploring fine-tuned versions of Clef for tasks such as evaluating Trust & Safety submissions, triaging support tickets, and distinguishing good bots from bad ones in its bot management products.

The fine-tuning pipeline draws on several existing Cloudflare platform components: AI Gateway to capture request and response data, Workers AI to generate model rollouts, Containers to serve as a sandbox for scoring and replaying agent actions, and a new Trainer component to update the fine-tuned model's weights before redeploying it through Workers AI's Bring Your Own Model capability, built on technology from Cloudflare's acquisition of Replicate.

Cloudflare frames the launch as an early but significant step in its broader AI platform strategy, describing decision models as a tool that could reshape how autonomous agents operate and positioning the company as infrastructure for what it calls the "agent cloud." The company says it is seeking design partners among existing customers with specific fine-tuning use cases.