Cloudflare releases open-source AI models for decision-making

Cloudflare introduced Clef, a set of open-weight AI models intended to choose between predefined options instead of generating free-form text. The company released 9B- and 27B-parameter versions, alongside a platform for adapting the models to specific decision-making tasks.

How Clef operates

Clef is a 27B multimodal model that accepts a state and a schema of typed questions as input. The model returns decisions, including a probability for each allowed option for every question in a single forward pass. It processes text, JSON, images, or video. Cloudflare’s API is compatible with the popular Typesafe AI’s Jev System One model.

The hyperscaler also released Clef-Flash, a smaller 9B multimodal model designed for latency-sensitive decisions. Benchmarks show it has a median latency of 38.8 ms in Cloudflare’s infrastructure, compared with 209.3 ms for the 27B Clef model.

Because the models are hosted on Cloudflare’s infrastructure, the company can take advantage of its GPUs at the edge. This arrangement leads to low network latency and faster decisions. Developers can place Clef in the hot path for agents to make decisions and combine it with one of Cloudflare’s LLMs on Workers AI to take action.

Clef differs from other decision models in a few ways. It has a vision encoder so it can take in images and classify visual content, which is not possible with Jev, which only does text classification today. The model also has a 64k context window, compared to Jev’s 32k, allowing more input state for the model to classify against.

Community response and future plans

On a popular Hacker News thread, the community discusses decision models, latency, and open weights. Jacek Złydach wrote that the technology is not a new paradigm, but a low-hanging fruit that has been available for years. SebastianSosa warned that public benchmarks are easy to cheat and suggested that releasing a public benchmark could distract developers from overfitting to the metric rather than making something actually useful.

Cloudflare plans to fine-tune Clef for specific use cases, such as support triage and bot classification, using historical labeled data. The company also announced a fine-tuning service that lets customers adapt Clef to their own workloads using their data. No firm date has been announced for the self-service platform.

The company has made the Clef models available through Workers AI and as downloadable weights on Hugging Face, inviting developers to experiment with them and provide feedback.