On October 1, 2026, Cloudflare unveiled two open-source models called Clef and Clef-Flash under the Apache-2.0 license. The models come in sizes of 27 billion and 9 billion parameters and feature native compatibility with interfaces established by TypeSafe AI and Jev. Rather than generating unpredictable conversational text like standard large language models, these decision models adopt a fundamentally different architecture. They process structured input states such as JSON data, text, and documents in a single forward pass, returning deterministic probability vectors across predefined action sets.
Bypassing token-by-token autoregressive generation delivers substantial improvements in execution speed. Clef-Flash completes decisions in roughly 38.8 milliseconds, while the larger 27-billion-parameter Clef model operates at approximately 209 milliseconds. This sub-second performance represents a marked improvement over conversational models that regularly require several seconds to return structured output. Enterprise workflows gain significant efficiency by eliminating token-generation overhead and stabilizing operational response times.
The practical advantages of this architecture are evident in standardized evaluation benchmarks for financial services. On the established BANKING77 benchmark, Clef achieved a macro-F1 score of 94.20 percent. This result surpasses the previous state-of-the-art model, Jev, which recorded a score of 79.74 percent. The benchmark assesses how accurately customer inquiries and transactional records are mapped to discrete banking intents and workflow queues.
In institutional finance and FinTech environments, decision models enable robust, hallucination-free gatekeeper architectures. Because the system outputs numerical probability distributions rather than generative language, the risk of confabulated information is eliminated by design. Financial institutions can deploy these systems for automated ticket routing, transaction risk scoring, and compliance triage. Human specialists are brought into the loop only when model confidence falls below strict, predetermined thresholds.
Cloudflare's release reflects a broader operational shift across institutional software engineering. After extensive experimentation with general-purpose generative agents, risk-conscious organizations are increasingly prioritizing bounded, deterministic systems for core infrastructure. Open decision models provide financial institutions with verifiable safety guarantees without locking them into proprietary third-party APIs. By making both models freely available under an Apache-2.0 license, Cloudflare provides a viable path toward predictable and cost-controlled automation.