On 1 October 2026, Cloudflare announced Clef and Clef-flash, the first models trained by the company’s Workers AI team. Unlike large language models that generate prose, a decision model reads an input state and a set of typed questions and returns a probability for every allowed answer. The agent gets a structured decision — route the ticket, block the request, escalate to a human — without parsing free text or waiting on reasoning tokens.
Clef, at about 27 billion parameters, targets highest-precision decisions; Clef-flash, at about 9 billion, targets millisecond latency on the hot path. Both offer a 64,000-token context window and are compatible with Typesafe’s Jev-style API, easing experiments and migrations. Cloudflare says Clef leads the Jev Decision Index at launch.
Weights ship on Hugging Face under Apache 2.0 for local runs, and hosted endpoints live on Workers AI as @cf/cloudflare/clef and @cf/cloudflare/clef-flash. Developers can call them via the env.AI.run() binding or REST at /ai/run, and wire AI Gateway to log traffic.
Cloudflare’s highlighted use cases are support triage, threat intelligence, trust and safety, and agent guardrails. In a domain-classification demo with Browser Run, Clef reportedly classified a site in 2.2 seconds versus 4.7 seconds for gpt-oss-120b in the same flow. For agents, the promise is checking “should I take this action?” in tens of milliseconds before calling a tool.
Alongside the models, Cloudflare debuted a reinforcement-learning fine-tuning product: capture data via AI Gateway, generate rollouts on Workers AI, score and replay actions in Containers, update weights with a Trainer, and redeploy the tuned model. In the first phase, fine-tuning is hands-on with FDE engineers; the goal is a self-serve platform.
The launch lands the same day Perplexity and AWS Strands Labs also released decision models, consolidating a new category: bill input only, free or local output, and optimise for agent routing rather than long essays. Capital & Compute counted ten new models on 1 October, none from OpenAI, Anthropic, Google, xAI, Meta or DeepSeek.
For Cloudflare, Clef fits the “agent cloud” thesis: the network that already terminates SSL, filters bots and runs Workers also wants to be where agents decide at the edge. A small typed model cuts cost and format hallucinations versus asking a giant LLM to “answer only yes or no.”
Risks remain. A poorly calibrated decision model can reject legitimate traffic or approve dangerous actions with high confidence. Probability calibration, policy rubrics and human review on the low-confidence tail are still essential. Apache 2.0 aids auditability but does not replace evaluation in the customer’s domain.
Competition with Typesafe’s Jev and with Perplexity and AWS deciders will turn on F1 quality, latency, price and fine-tuning ease. Week-of coverage cited Clef around 94 F1 versus ~80 for Jev on a public benchmark, with substantially lower latency; those figures must be verified on each company’s real loads.
Enterprises already using Cloudflare for WAF, Bot Management or Workers have a short path to insert Clef into existing pipelines. Teams living in other clouds must weigh Workers AI lock-in against open weights that can run anywhere.
Neural Pulse sees in Clef the maturation of a simple idea: many agent steps do not need an essay; they need a distribution over a closed set of actions. On 1 October 2026 that idea left the niche-paper stage and became a product from a global infrastructure provider.
We will watch pricing, the self-serve RL roadmap, and whether frontier labs answer with their own deciders or keep betting that one generative model can do it all.
Integrators should version decision rubrics as code: every policy change is an effective model change. Without that, RL fine-tuning will optimise a ghost objective.
Jev compatibility eases parallel Clef vs Jev vs Perplexity tests. A single harness that speaks all three cuts switching costs when price or latency moves.
In regulated sectors (banking, health), documenting why a probability threshold was chosen will matter as much as the threshold itself to auditors.
