Cloudflare has released Clef and Clef-flash, two open-weight decision models designed to make fast, deterministic classifications for autonomous agent workflows. Unlike large language models that generate open-ended text, decision models produce bounded, structured outputs that can route tasks, trigger escalations, or defer to humans based on probability scores.
According to Cloudflare, decision models differ from LLMs by providing consistent, typed outputs without constant retraining when classification categories change. The company is releasing both models on Hugging Face under an Apache 2.0 license and hosting them on Workers AI infrastructure.
Clef currently leads the Jev Decision Index benchmark. The models have distinct advantages: Clef includes a vision encoder for image classification (unlike Typesafe AI’s Jev, which handles text only) and supports a 64k context window compared to Jev’s 32k. Across 43 evaluation benchmarks, Clef and Clef-flash achieved faster median latency than competing models—209.3ms and 38.8ms respectively, while Jev took 524.1ms.
Cloudflare demonstrated the models on its own Threat Intelligence team, using Clef to classify website domains. Processing a domain through the model—fetching, rendering, and classifying it—took 2.2 seconds, compared to 4.7 seconds for the company’s fastest LLM, gpt-oss-120b. Clef also returned more classifications.
On Typesafe’s benchmark suite, Clef beat Jev in 3 of 4 workflow areas: invoice processing (64.7 vs 57.1), customer service (76.3 vs 77.0), and agent trace observability (68.5 vs 69.8). The security incidents category showed comparable scores (62.9 vs 61.7).
Cloudflare is also introducing a reinforcement learning fine-tuning platform that allows customers to optimize Clef for specific use cases. The company guarantees it does not read, store, or train on customer requests unless customers opt into the fine-tuning product.
The models are Jev API-compatible, allowing users to swap them into existing workflows with minimal changes. Developers can access them through Cloudflare’s API or experiment locally with the open-source versions on Hugging Face.
Key facts
- Cloudflare released Clef and Clef-flash decision models under Apache 2.0 license on Hugging Face
- Clef scores highest on the Jev Decision Index benchmark for decision models
- Clef includes vision encoding for image classification and supports 64k context window, both advantages over Jev
- Clef processed domain classifications 2.2x faster than Cloudflare’s fastest LLM in threat intelligence testing
- Models achieve median latency of 209.3ms (Clef) and 38.8ms (Clef-flash) across 43 benchmarks
- Cloudflare introduced new RL fine-tuning platform for customizing Clef to specific use cases
- Models are hosted on Cloudflare Workers AI infrastructure at the edge for low network latency