Clef: Cloudflare Launches Its Own Open Source Decision Models

A new category: the decision model

Cloudflare starts from an observation: over the past few weeks, Jev System One, Typesafe AI's model, has popularized the idea of a "decision model." It's not an LLM that reasons and generates text, but a model that produces structured, bounded outputs, quickly, cheaply, and consistently.

The principle: you provide a state (a support ticket, for example) and typed questions. The model returns answers with probabilities: is it urgent, which team should handle it, what's the severity. The code can then route, escalate, or hand off to a human.

Unlike classic classifiers, these models don't need to be retrained every time a new category appears: the options are defined in the request itself. Cloudflare sees this as a building block for agents that decide and act without a human in the loop, except when they judge it necessary.

The internal use case: classifying domains

Cloudflare's Threat Intelligence team tested Clef, paired with Browser Run, to categorize domain names. Example given: 95% fashion, 85% e-commerce, under 1% phishing.

The whole pipeline (fetching, rendering, and classifying the page) takes 2.2 seconds with Clef. With gpt-oss-120b, Cloudflare's fastest general-purpose LLM, the same workflow takes 4.7 seconds and returns only two categories.

The name itself comes from music: the clef placed at the start of a staff fixes the names of the notes that follow, just as the model fixes the framework for upcoming actions. And "CF" nods to Cloudflare.

What Clef brings compared to Jev

Three arguments are put forward. First, vision: Clef can classify visual content. Second, context: 64k tokens versus 32k for Jev. Finally, quality, measured on benchmarks drawn from the Jev Decision Index.

The gaps are sometimes striking. On BANKING77, Clef scores 94.20 macro-F1 versus 79.74 for Jev; on CLINC150+OOS, 97.43 versus 89.27; on the "home appliances" task, Clef-flash reaches 97.73 versus 52.27 for Jev.

But Jev stays ahead on When2Call (80.97 vs 72.37) and BRIGHT (47.52 vs 45.91). Clef-flash also collapses on CLINC150+OOS (66.77). On PhishNChips, another model in the table beats Clef (85.35 vs 79.60).

On latency, Clef-flash posts 38.8 ms median and 122.4 ms p95. Only Laya is faster (5.8 ms median), but its quality scores are very low. Edge GPU hosting also cuts network latency, which lets you place Clef in an agent's hot path.

Under the hood: no generation, just scoring

Clef builds on an experiment Cloudflare published the same week Jev launched: adapting DiffusionGemma to extract deterministic probabilities via logprobs, drawing on work by Matt Mastracci, a vLLM contributor.

The backbone is now Qwen, frozen: Qwen3.8-27B for Clef, Qwen3.5-9B for Clef-flash. Cloudflare trains a routing head and rank-256 LoRA adapters. At inference time, a single prefill pass is enough, then the schema's valid choices are scored in parallel, with no token-by-token text generation.

The architecture relies on a two-stage attention routing: each option retrieves the relevant parts of the prompt, then the fields attend to each other and re-read the payload before scoring. Training combines cross-entropy with label smoothing and a Brier loss to calibrate probabilities, on synthetic data that permutes field order, prompts, and schemas.

Added to this is RLCD (Reinforcement Learning for Calibrated Decisions): partial credit for nearby ordinal choices, reward for fully correct records, and a reference penalty to prevent distribution drift.

The real bet: RL fine-tuning

Cloudflare's internal teams want specialized versions of Clef: triaging Trust & Safety reports, sorting support requests, distinguishing good from bad crawlers in Bot products. Cloudflare says it has over 15 years of network data and labeled decisions to draw on.

The service starts with the FDE (forward-deployed engineers) team, in a white-glove mode. The goal is then a self-serve platform to capture data, fine-tune, and redeploy, entirely within Cloudflare.

The pipeline relies on existing building blocks: AI Gateway to automatically build datasets from traffic, Workers AI to generate rollouts, Containers as an RL sandbox, a new Trainer component to update weights, and Bring Your Own Model (Cog, inherited from the Replicate acquisition) to redeploy.