Clef classifies a domain in 2.2 seconds, and its weights are public
Cloudflare's open-weight Clef decision model classifies a domain in 2.2 seconds, and open weights hand its evaluation function to whoever you're classifying.
Cloudflare’s new Clef model took a website domain, fetched and rendered the page, and classified it in 2.2 seconds, returning probabilities like 95% fashion, 85% ecommerce, and under 1% phishing. The fastest general-purpose LLM Cloudflare ran through the same workflow, gpt-oss-120b, took 4.7 seconds and came back with two labels. That gap is the pitch for Clef and Clef-flash, two decision models now on Workers AI and open-sourced on Hugging Face under Apache 2.0.
A decision model is a classifier with a fixed job. You hand it a block of state and a set of typed questions, and it returns typed answers with probabilities instead of prose. The API example classifies a support message: is it urgent, which team should handle it (a choice among billing, technical, and sales), and how severe the impact is on an ordinal scale from no impact to critical. Your code reads the probabilities and routes the ticket, escalates, or defers to a human. The category comes from Typesafe AI’s Jev System One, and Clef is deliberately Jev-API compatible so you can swap one for the other. Cloudflare says Clef leads the Jev Decision Index, adds a vision encoder that Jev lacks, and carries a 64k context window against Jev’s 32k.
The speed comes from not generating text. Clef uses a Qwen backbone for a prefill-only pass, then scores the valid schema choices in parallel. The decision step is non-autoregressive, so there is no token-by-token generation and no intermediate string to parse. Cloudflare froze Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash and trained a routing head alongside rank-256 low-rank adapters, using label-smoothed cross-entropy for the schema outputs and a Brier loss to calibrate the probabilities. They also describe a reinforcement-learning objective, RLCD, that gives partial credit for adjacent ordinal choices and penalises drift from the reference model. The lineage runs back to an earlier experiment adapting DiffusionGemma to emit deterministic probabilities by exposing logprobs, built on Matt Mastracci’s vLLM work.
The output is built to be acted on without a human reading it. Cloudflare is explicit that a person no longer has to be in the loop: the agent gathers context, the decision model classifies, and the agent takes the action. That makes calibration the load-bearing property, which is why the Brier loss and the partial-credit reward exist. A 95% that should have read 60% becomes an auto-escalation, a refund, or a block before anyone sees it.
Cloudflare has been testing Clef on its own Threat Intelligence team to classify website domains, including a phishing probability, and the stated fine-tuning targets include Trust & Safety triage and deciding whether a crawler is a good bot or a bad bot. Fine-tuning is sold as a service: the forward-deployed engineering team tunes Clef on your labelled data first, with a self-serve platform to follow that wires AI Gateway to capture request/response pairs into a dataset, Workers AI to generate rollouts, Containers as an RL sandbox, a new Trainer to update weights, and bring-your-own-model deployment back onto Workers AI.
The weights are open, so anyone can pull Clef from Hugging Face and run it locally, which cuts both ways for a classifier that decides good bot from bad bot. If your abuse detection comes down to a probability from a model whose weights are public, whoever is probing you can run the same model offline and tune their inputs until the score lands where they want. Open weights are the right call for inspection and portability, and they also hand the evaluation function to the other side. A decision model cheap enough to drop into your hot path is cheap enough for the traffic hitting that path.
If you adopt one, treat the probability as an input to a policy rather than the policy itself. Keep a defer path for the low-confidence band, log the model’s inputs and outputs so a bad decision is reconstructable, and decide up front which actions a classifier may take unattended and which still need a person. Cloudflare’s guarantee that it doesn’t read, store, or train on your requests or responses carries one exception: the fine-tuning product, where AI Gateway captures your traffic so you can build a dataset from your own request and response data.
Keep Reading
supply chain700 French supermarkets used the date as a password
Date-rotating default passwords froze French supermarkets, backdoored Ledger wallets stole $100M, Pwn2Own hit 98 zero-days, and OpenAI agents crossed lines.
AWSAgentCore and the Metadata Problem
Zenity's AgentCorruption research shows how one prompt to an AWS Bedrock AgentCore agent could expose credentials and widen blast radius.
AI agentsClaude optimizes what it measures
How Anthropic made claude.ai about 3x faster in two weeks by turning every metric an internal Claude model could measure into a CI benchmark it could climb.
Latest on the Wire
Full wire →- AI Coding Agents Boost Output But Code Review Bottlenecks IncreaseHacker News
- AI Overviews Cut Organic Traffic by 40%Hacker News
- Apache Patches Critical Roller RCE VulnerabilityBleepingComputer
- Complete Web Browser Built from Scratch in PythonHacker News
New signal daily · RSS
Stay in the loop
New writing delivered when it's ready. No schedule, no spam.