Cloudflare open-sources Clef, a decision model that returns typed probabilities instead of text

To understand the launch, start with a counterintuitive definition: a decision model does not write text. Cloudflare's own example — you pass in a customer support message (inputs) and ask whether it is urgent and which team should handle it; the model returns typed answers with probabilities (outputs) that your code uses to route the ticket, trigger an escalation or defer to a human. Cloudflare makes the implication explicit: a human no longer necessarily needs to be in the loop for agentic decisions, and agents can gather context, decide and act programmatically, deferring to a human only when needed. That is the boundary with general LLMs — the LLM handles open-ended reasoning and generation, while the decision model makes the 'pick one' step cheap, fast and consistent without retraining every time a category is added.

Cloudflare's measured example comes from its own Threat Intelligence team classifying website domains. Given a domain (with Browser Run), Clef quickly identifies the categories it falls under — Cloudflare's illustration being a 95% chance it is a fashion site, 85% ecommerce, under 1% phishing. The comparison Cloudflare reports: Clef took 2.2 seconds to fetch, render and classify, while its fastest general LLM, gpt-oss-120b, took 4.7 seconds in the same workflow and returned only two classifications. Cloudflare extrapolates that combination of halved latency and fuller results to any use case needing quick programmatic decisions, describing it as unlocking agentic workflows that can autonomously decide, reason and execute.

Cloudflare lists three differentiators. First, Clef has a vision encoder, so it can take in images and classify visual content — Cloudflare says Jev is text-only today. Second, a 64k context window versus what Cloudflare says is Jev's 32k, letting users squeeze in more input state to classify against. Third, competitive scores across quality benchmarks; Cloudflare publishes the full table covering BFCL, ToolRet, API-Bank, Home appliances, When2Call, BANKING77, CLINC150+OOS, BRIGHT, Amazon ESCI and PhishNChips, and says that across the 43 eval benchmarks it ran, Clef models generally beat comparable decision models on latency — the one exception being Laya, which Cloudflare describes as very fast but trading off quality. On the latency table, the medians Cloudflare reports are Clef 209.3 ms, Clef-flash 38.8 ms, Jev 524.1 ms, DiffusionGemma Jev 84.4 ms, Kev-9B 51.4 ms and Laya 5.8 ms; p95 is Clef 238.6 ms, Clef-flash 122.4 ms and Jev 536.0 ms.

Cloudflare also goes deep on training: the backbone is a Qwen model, with Qwen3.8-27B frozen for Clef and Qwen3.5-9B for Clef-flash, jointly optimizing a routing head alongside rank-256 low-rank adapters. At inference, Clef runs a prefill-only pass through Qwen, then scores all valid schema choices in parallel; the decision step is non-autoregressive, so there is no intermediate text generated token by token, making it significantly faster than autoregressive LLMs. Training uses label-smoothed cross-entropy paired with a Brier loss to calibrate probabilities, plus a reinforcement-learning objective called RLCD (Reinforcement Learning for Calibrated Decisions). On deployment, Cloudflare says the models are hosted on Workers AI and run on its edge GPUs, fit for the agent hot path, and that enterprise customers get the guarantee that Cloudflare does not read, store or train on requests or responses unless they opt into fine-tuning. Fine-tuning runs through AI Gateway to capture data, Workers AI for rollouts, Containers as the RL sandbox, a new Trainer to update weights, and redeployment back onto Workers AI.