Cloudflare has officially entered the AI model training arena, marking a significant shift in its identity from a pure network infrastructure provider to a developer platform capable of hosting in-house intelligence. On October 1, 2026, the company unveiled Clef and Clef-flash, its first internally trained open-weight models. These tools are now live on the Workers AI platform, designed specifically to handle fast, structured decision-making processes that are essential for modern agentic workflows.
Two Models, Two Speeds
The new release features a dual-model strategy tailored to different performance requirements. Clef serves as the heavyweight option, packing 27 billion parameters and built upon the Qwen3.8-27B foundation. It offers a substantial 64k-token context window and supports diverse input types, including text, JSON, images, and video. This makes it suitable for complex tasks that require deep contextual understanding.
In contrast, Clef-flash is the lighter sibling, featuring 9 billion parameters based on the Qwen3.5-9B architecture. Cloudflare has positioned Clef-flash for latency-critical paths within applications. In these scenarios, even a fraction of a second of delay is unacceptable. Both models include a vision encoder capable of processing up to four images per request, ensuring that visual data can be integrated into decision logic without significant overhead.
Performance and Pricing
Cloudflare emphasizes that these models are optimized for outputs that are fast, accurate, and deterministic. To demonstrate this, the company highlighted results from the BANKING77 classification benchmark. Clef achieved a macro-F1 score of 94.20, significantly outperforming Jev, a product from Typesafe, which scored 79.74. The speed advantage is even more pronounced with the lighter model. Clef-flash posted a median latency of 38.8 milliseconds, whereas Jev recorded a median latency of 524.1 milliseconds.
The pricing structure reflects this efficiency. Clef is priced at $0.24 per million input tokens, while Clef-flash costs $0.09 per million input tokens. Notably, output tokens are not billed at all, a feature that could substantially reduce costs for developers running high-volume inference tasks. Furthermore, both models are compatible with the Jev-API. This compatibility allows teams that have already built integrations around Jev to swap in Cloudflare’s models without rewriting their existing code, lowering the barrier to entry for adoption.
The weights are fully open-sourced on Hugging Face under the Apache 2.0 license.
Edge Infrastructure and Developer Tools
A critical component of this launch is the infrastructure supporting these models. Clef and Clef-flash run on Cloudflare’s edge GPU resources. By utilizing servers located close to end users, the platform reduces the distance data must travel, which is crucial for maintaining low latency. This approach aligns with Cloudflare’s long-standing focus on global network performance, now extended into the realm of AI inference.
Alongside the model release, Cloudflare is introducing a reinforcement learning fine-tuning platform. The company states that this platform will launch with hands-on support for customers, allowing developers to customize the models for specific use cases. This addition transforms Workers AI from a simple hosting environment into a comprehensive developer platform for building and refining AI agents.
Strategic Implications for Developers
The launch occurred during Cloudflare’s Birthday Week, its annual period of product announcements. While Workers AI has hosted third-party models in the past, Clef and Clef-flash represent the first models trained in-house by Cloudflare. This move signals a strategic deepening of the company’s AI capabilities. For developers, the practical impact is immediate: access to high-performance, open-weight decision models that are natively integrated into a global edge network.
The open-source nature of the weights, released under the Apache 2.0 license, encourages community scrutiny and customization. Developers can inspect the model architecture, fine-tune it for niche applications, and deploy it with confidence in its transparency. This transparency is particularly valuable in sectors where deterministic behavior and auditability are required, such as financial services or automated compliance workflows.
What to Watch Next
As Cloudflare expands its AI portfolio, readers should monitor the rollout of the reinforcement learning fine-tuning platform. The success of this tool will determine how easily developers can adapt Clef and Clef-flash to their specific business logic. Additionally, the competitive landscape may shift as other infrastructure providers respond to Cloudflare’s aggressive pricing and performance metrics. The ability to swap models via the Jev-API compatibility suggests a future where model choice is fluid, driven by performance and cost rather than vendor lock-in. For technical teams, the immediate next step is to evaluate these models against their current decision-making pipelines to assess potential latency and cost savings.