Cloudflare Unveils Clef and Clef-flash: Precision AI with Typed Probabilities

A New Era for AI Decision-Making

The central development is this: Cloudflare has introduced a groundbreaking advancement in artificial intelligence with the release of Clef and Clef-flash. These innovative models, the first to be developed by Cloudflare’s Workers AI team, represent a departure from traditional large language models (LLMs) by focusing on precise, typed probabilities rather than generating free-form text. This shift promises to streamline AI integration and enhance the reliability of automated decision-making across various applications.

Meanwhile, Unlike generative AI chatbots, Clef and Clef-flash are designed as dedicated decision models. They ingest an input state alongside a structured schema of specific questions and, in return, provide a probability for every permissible answer. This eliminates the need for complex parsing of free-form text outputs, offering a more direct and actionable form of AI intelligence.

Understanding Decision Models: Beyond Chatbots

Traditional LLMs are adept at generating human-like text, but their output often requires additional processing to extract structured data or make specific decisions. Decision models, like Clef and Clef-flash, tackle this challenge head-on by providing direct, quantifiable answers to a predefined set of questions.

Cloudflare’s new models support three primary question types:

• Yes/No (Noul): Returns a probability indicating the likelihood of a ‘yes’ answer.
• Choice: Selects one named option from a list, providing probabilities for each option and an overall confidence score.
• Score: Rates an input against an ordered rubric, delivering a probability-weighted score.

These capabilities make Clef and Clef-flash ideal for tasks requiring clear, unambiguous answers, such as classification, risk assessment, and structured data extraction. Furthermore, they are open-weight under the Apache 2.0 license and fully compatible with TypeSafe AI’s Jev API, fostering broader adoption and integration within the AI ecosystem.

Deployability and Technical Foundations

For example, Both Clef and Clef-flash are immediately deployable on Cloudflare’s Workers AI platform, offering seamless integration for developers. For those preferring self-hosting, the model weights are readily available on Hugging Face.

Under the hood, Clef is post-trained from Qwen3.8-27B, while Clef-flash leverages Qwen3.5-9B. Both retain their backbone’s vision encoder, allowing them to process not only text but also images. On Workers AI, a single request can handle up to 64 questions and up to 4 images, showcasing their robust multi-modal capabilities.

How Clef Processes Information

The inference process for Clef models involves two main stages:

1. Backbone Prefill: The primary backbone runs a single pass over the input state and questions.
2. Joint Schema Head: A smaller transformer then reads the final hidden states, routing evidence to each question, allowing fields to interact, and jointly scoring all options. A per-question softmax function converts these scores into precise probabilities.

Training involved freezing the backbones and optimizing the routing head using rank-256 low-rank adapters. The loss function combined label-smoothed cross-entropy with a Brier loss for calibration, alongside a secondary objective, Reinforcement Learning for Calibrated Decisions (RLCD), which provides partial credit for adjacent ordinal choices.

Performance Benchmarks: Speed and Accuracy

Interestingly, Cloudflare’s internal benchmarks, using a 10-benchmark shortlist from the Decision Index 0.2.1 suite, reveal impressive performance from Clef models. They outperformed competitors, including Jev, on 7 out of 10 tests:

• BANKING77 (macro-F1): Clef scored 94.20% compared to Jev’s 79.74%.
• CLINC150+OOS (macro-F1): Clef achieved 97.43% against Jev’s 89.27%.
• Home Appliances (case exact): Clef-flash demonstrated strong performance at 97.73% versus Jev’s 52.27%.

While Jev maintains leads on knowledge-intensive tests like GPQA Diamond, MMLU-Pro, and BBH, Clef models show significant advantages in specific workflow evaluations. For instance, Clef surpassed Jev in 3 out of 4 TypeSafe workflow areas, including invoice processing (64.7% vs. 61.8%), customer service (76.3% vs.

76.0%), and security incidents (62.9% vs. 61.7%).

However, Speed is another critical advantage. In Cloudflare’s threat intelligence workflow, Clef classified a domain in a mere 2.2 seconds, significantly faster than gpt-oss-120b, which took 4.7 seconds. This efficiency translates directly into faster, more responsive AI applications.

Flexible Deployment and Customization

Clef and Clef-flash offer versatile deployment options. Developers can call the models via the Workers AI binding (env.AI.run()), the REST API, or through AI Gateway. For organizations requiring on-premise solutions, the model cards confirm successful testing for self-hosting on a single H200 GPU with BF16 weights.

Meanwhile, Cloudflare is also rolling out a reinforcement learning service specifically for fine-tuning Clef models with private data. Initially available through Cloudflare’s forward-deployed engineers, a self-serve platform is planned for the future. This service combines AI Gateway, Workers AI, Containers, and a new Trainer component, empowering teams to tailor Clef’s decision-making capabilities to their unique operational needs.

The Future of Precision AI

Cloudflare’s Clef and Clef-flash mark an important step towards more precise, efficient, and deployable AI. By delivering typed probabilities instead of ambiguous text, these open-weight decision models offer a clear path to building more reliable and impactful AI applications, setting a new standard for how we interact with and leverage artificial intelligence for critical decisions.

Expert Perspective

From an industry angle, the clearest signal around Cloudflare Clef Decision Models is how it may influence clef. The story reads less like a one-day spike and more like a marker of broader movement.

The next phase will depend on how quickly teams, regulators, or customers react. In practice, that gives Cloudflare Clef Decision Models room to reshape expectations across models over the near term.

For readers focused on practical impact, the best next step is to watch what changes around cloudflare once attention turns into execution.

Frequently Asked Questions

Why does Cloudflare Clef Decision Models matter right now?

A New Era for AI Decision-MakingThe central development is this: Cloudflare has introduced a groundbreaking advancement in artificial intelligence with the release of Clef and Clef-flash.

What broader change could Cloudflare Clef Decision Models signal?

These innovative models, the first to be developed by Cloudflare’s Workers AI team, represent a departure from traditional large language models (LLMs) by focusing on precise, typed probabilities rather than generating free-form text.

What should the market watch next around Cloudflare Clef Decision Models?

This shift promises to streamline AI integration and enhance the reliability of automated decision-making across various applications.Meanwhile, Unlike generative AI chatbots, Clef and Clef-flash are designed as dedicated decision models.

Source: https://www.marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/