The company made both models available through Workers AI on October 1 and released their weights on Hugging Face under the Apache 2.0 licence. Clef has 27 billion parameters, while the latency-focused Clef-flash has 9 billion. Both accept a state and typed questions and return probabilities for permitted answers rather than generating free-form prose.
Cloudflare said the models are fully compatible with the Jev and System One application programming interface, allowing developers with existing Jev integrations to switch models largely by changing the endpoint and model identifier. The models support yes-or-no, multiple-choice and ordered-score questions, and can handle as many as 64 questions in a request.
The launch puts Cloudflare directly alongside TypeSafe AI, which introduced Jev on September 15 as its first public System One model. TypeSafe designed Jev for software automation where applications need bounded, probabilistic decisions rather than conversational output. Its interface similarly takes unstructured state information and structured questions and returns typed answers that software can use immediately.
Cloudflare is claiming an advantage on both speed and benchmark performance, although the comparative results were produced by Cloudflare and should not be treated as independent validation. Across 43 benchmark runs, the company measured median latency of 209.3 milliseconds for Clef and 38.8 milliseconds for Clef-flash, against 524.1 milliseconds for Jev. It said that made Clef about 2.5 times faster at the median and Clef-flash about 13 times faster.
Its published evaluation also showed a Clef model leading Jev on most of 10 decision benchmarks. Clef recorded 98.47 on the BFCL case-exact test against Jev’s 95.75, while its BANKING77 macro-F1 score was 94.20 compared with 79.74 for Jev. Jev remained ahead on some tasks, including the When2Call and BRIGHT evaluations, underscoring that the results do not establish a universal performance lead.
Cloudflare also tested the models against workflow evaluations developed by TypeSafe. It said Clef outperformed Jev in three of four areas — invoice processing, customer service and security incidents — while Jev led on agent-trace observability. Those comparisons likewise come from Cloudflare’s own testing.
The architecture reflects the narrower role assigned to decision models. Rather than producing text token by token, Clef scores allowed options in a single forward pass. Cloudflare says this avoids output parsing and reduces the latency associated with reasoning and text generation, making the models suitable for operations such as routing customer-support requests, identifying security incidents or deciding whether an automated process should escalate to a person.
Clef is post-trained from Qwen3.8-27B and retains its vision encoder, while Clef-flash is post-trained from Qwen3.5-9B. The released model cards describe both as multimodal, able to process text, JSON, images or video alongside a schema of questions. Their weights can be downloaded and run outside Cloudflare’s hosted service, a distinction that could matter to developers seeking greater control over deployment.
Cloudflare is coupling the model release with a reinforcement-learning fine-tuning service. The company said design partners will be able to adapt Clef to their own workloads, using components of its AI platform to capture traffic, generate training rollouts, score actions in sandboxed environments and redeploy tuned models.
The move broadens competition around a category TypeSafe describes as System One models, inspired by the concept of rapid, intuitive decision-making. TypeSafe says Jev is intended to supply probabilities and confidence for decisions that applications can consume directly, rather than imitate the open-ended behaviour of a general-purpose chatbot.
Cloudflare’s implementation preserves that basic interface while extending inputs to visual material and offering a 65,536-token context window. TypeSafe’s public API documentation lists Jev as a general-purpose System One model and exposes a systemone endpoint for submitting state, model and question fields.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
