Cloudflare Introduces Clef and Clef-flash
Cloudflare has officially released Clef and Clef-flash, marking a significant shift in how artificial intelligence is integrated into web infrastructure. These are the first models trained by Cloudflare's Workers AI team, designed specifically as decision models rather than traditional chatbots. By moving away from free-form text generation, Cloudflare aims to provide developers with a more deterministic and efficient tool for processing structured data within web applications.
Defining the Decision Model Approach
Unlike Large Language Models (LLMs) that generate tokens sequentially and often require complex parsing of unstructured output, Clef operates on a fixed schema. Each model reads an input state and a set of typed questions, returning a probability score for every allowed answer. This approach eliminates the ambiguity of natural language generation, making it ideal for automated workflows where precision and speed are critical. Both models are released under the Apache 2.0 license, ensuring they are open-weight and accessible for self-hosting via Hugging Face.
Clef supports 3 question types: On Workers AI, 1 request carries up to 64 questions and up to 4 images.
Compatibility with the System One API
A key feature of Clef is its compatibility with TypeSafe AI’s Jev API. TypeSafe AI launched Jev, its first 'System One' model, in September 2026, followed by open alternatives like Kev-9B and Laya. Because Clef uses the same System One API, developers can switch from Jev to Clef by simply changing the endpoint and model name. This standardization lowers the barrier to entry for teams already invested in the System One ecosystem, allowing for seamless integration without rewriting core logic.
Technical Architecture and Training
Under the hood, Clef is post-trained from Qwen3.8-27B, while Clef-flash is derived from Qwen3.5-9B. Both models retain the backbone’s vision encoder, enabling multimodal processing. The inference process is divided into two stages: first, the backbone performs a single prefill-only pass over the state and questions. Second, a small transformer known as the joint schema head reads the final hidden states. This head routes evidence to each question, allows fields to cross-attend, and scores all options jointly. A per-question softmax then converts logits into probabilities.
During training, the backbones were frozen, and the routing head was jointly optimized using rank-256 low-rank adapters. The loss function combines label-smoothed cross-entropy with a Brier loss for calibration. Additionally, a secondary objective called Reinforcement Learning for Calibrated Decisions (RLCD) provides partial credit to adjacent ordinal choices, enhancing the model's ability to handle nuanced decision-making.
Performance Benchmarks
According to Cloudflare’s internal Decision Index run, Clef demonstrated strong performance across a 10-benchmark shortlist from the Decision Index 0.2.1 suite. Clef scored highest on seven of these benchmarks. However, Jev maintains clear leads in specific areas, such as GPQA Diamond (78.3 vs. 48.0) and MMLU-Pro (82.7 vs. 65.9). In TypeSafe’s workflow evaluations, Clef outperformed Jev in three out of four areas, including invoice processing and customer service, with only slight margins. Notably, in a threat intelligence workflow, Clef classified a domain in 2.2 seconds, compared to 4.7 seconds for gpt-oss-120b. It is important to note that all these numbers are vendor-reported, and independent replication is yet to be conducted.
Deployment and Developer Tools
For developers, both models are callable through the Workers AI binding (env.AI.run()), the REST API, or AI Gateway. For those preferring self-hosting, the model cards indicate testing on a single H200 GPU with BF16 weights. Cloudflare has also announced a reinforcement learning service that allows teams to tune Clef on private data. This pipeline integrates AI Gateway, Workers AI, Containers, and a new Trainer component. Initially, this service will be available to Cloudflare’s forward-deployed engineers, with a self-serve platform planned for later release. Teams interested in early access can apply via the design partner form.
Conclusion
The release of Clef and Clef-flash represents a strategic move by Cloudflare to empower developers with AI tools that are faster, more predictable, and easier to integrate into existing web architectures. By focusing on decision-making rather than conversation, Cloudflare addresses a specific pain point in enterprise AI adoption. As the ecosystem around the System One API grows, developers should watch for the rollout of the self-serve tuning platform and independent benchmarks that validate the vendor-reported performance metrics.