--- title: "Cloudflare's Clef and Clef Flash: AI Models That Answer in Probabilities, Not Prose" slug: cloudflare-clef-decision-models-2026 category: ai category_label: "AI" author: "BrainWavePost Staff" date: 2026-10-03 tags: ["Cloudflare", "Clef", "Clef Flash", "decision models", "Workers AI", "AI agents", "open weights", "edge computing"] read_time_minutes: 8 canonical_url: https://brainwavepost.com/article/cloudflare-clef-decision-models-2026 source: BrainWavePost --- # Cloudflare's Clef and Clef Flash: AI Models That Answer in Probabilities, Not Prose *AI · 2026-10-03 · BrainWavePost Staff · 8 min read* > Cloudflare has released Clef and Clef Flash, open-weight 'decision models' that return typed answers with confidence scores instead of free text. They're fast, multimodal and Apache-licensed — and pitched as letting AI agents act on more decisions without a human checking each one. Here's a plain-English, sourced explainer. > **The one-line version** _(info)_ > > A normal language model replies with text you then have to parse. A 'decision model' replies with a structured answer — yes/no, a chosen option, or a rating — plus a confidence score. Cloudflare's Clef and Clef Flash do exactly that, fast and at the edge, which is useful for routing and classification inside automated systems. [1][2] On October 1, 2026, Cloudflare released two open-weight models called Clef and Clef Flash, along with a new reinforcement-learning fine-tuning platform. [1][2] They belong to a less flashy but increasingly important category: 'decision models' — models built not to write paragraphs but to make small, bounded, repeatable choices with a confidence number attached. [1][3] ## What a 'decision model' actually is Large language models are optimized to generate text. That's powerful, but awkward when all you need is a decision: is this support ticket urgent? Is this request a bot? Which of five queues should this go to? Parsing those answers out of free-form text is error-prone, and the model can drift in how it phrases things. [2][3] A decision model instead returns a typed, bounded result plus a probability. Clef and Clef Flash handle three question shapes: yes/no, pick one named option, and rate on a scale — and they can take up to 64 questions in a single request. [2] Because the output is structured and comes with a confidence score, downstream code can act on it directly: route the request, trigger an action, or escalate when confidence is low. [2][3] - **2 models** — Clef (~27B params) and Clef Flash (~9B), both built on open Qwen bases [2] - **~39 ms** — Clef Flash median latency Cloudflare reports (≈209 ms for Clef) [2] - **64K / 64** — token context, and up to 64 questions per request [2] - **Apache 2.0** — open-weight license, published on Hugging Face [1][2] ## What's actually new here - Multimodal decisions: unlike the incumbent decision model (TypeSafe's 'Jev'), Clef and Clef Flash can read images and video frames alongside text. [2] - Speed at the edge: hosted on Cloudflare's Workers AI, inference runs close to the request. Cloudflare reports median latencies around 39 ms (Clef Flash) and 209 ms (Clef) versus about 524 ms for Jev — vendor figures, so worth independent testing. [2] - Open weights: both ship under Apache 2.0 on Hugging Face, so teams can run or adapt them, not just call an API. [1][2] - Drop-in compatibility: Cloudflare says the models are Jev-API compatible, lowering the switching cost for teams already using that pattern. [2] - An RL fine-tuning platform: customers can log requests through Cloudflare's AI Gateway, build a dataset, evaluate in a container-based RL sandbox, and deploy a fine-tuned model back on Workers AI. [1][2] ## Where it's meant to be used Cloudflare says its own teams use Clef-style models to triage support requests, detect bad bots, and assess Trust & Safety submissions — the kind of high-volume, repetitive judgment calls that sit in the 'hot path' of a service. [1][2] The pitch is that putting a fast, calibrated decision model inline lets automated workflows classify and route without a slow round-trip to a general-purpose LLM. [2][3] ## The 'humans out of the loop' claim — read it carefully > **What confidence scores do and don't change** _(note)_ > > Cloudflare framed Clef as reducing the need for a human in the loop on agent decisions. The sensible version of that is narrow: because each answer comes with a calibrated probability, a system can act automatically when confidence is high and escalate to a person when it isn't. That can cut routine human checks — it does not mean oversight disappears, and a confident wrong answer is still wrong. [2][3] The honest framing is that decision models change where human judgment is spent, not whether it's needed. Confidence thresholds, auditing of automated decisions, and review of the low-confidence tail still matter — arguably more, once more decisions are automated. Treat 'no human needed' as a marketing simplification of a real but bounded efficiency gain. [3] ## What to keep in mind - The latency and accuracy numbers are Cloudflare's own; benchmark them on your workload before relying on them. [2] - A typed probability is only as good as the model's calibration — a high score is not a guarantee of correctness. [3] - Open weights (Apache 2.0) are a genuine plus for portability and self-hosting, separate from the hosted Workers AI offering. [1][2] - 'Decision model' is a design pattern, not magic: it narrows what the model outputs, which reduces parsing errors but doesn't eliminate the need to validate automated actions. [3] The honest summary: Clef and Clef Flash are a credible, well-targeted release — fast, multimodal, open-weight decision models with a fine-tuning path and edge hosting, aimed at the unglamorous routing-and-classification work inside modern AI systems. The capability is real; the 'humans no longer needed' framing is the part to read skeptically. [1][2][3] > **Editorial note** _(tip)_ > > This is technology journalism, not an endorsement or purchasing advice. Specifications, benchmark figures and availability reflect Cloudflare's announcement and early reporting as of publication; vendor-reported performance should be independently verified for your use case. ## Sources and further reading - [1] Cloudflare — 'Introducing Clef: our open-source decision models, and new RL fine-tuning platform' (official blog): https://blog.cloudflare.com/clef-decision-models/ - [2] MarkTechPost — 'Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text' (2026): https://www.marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/ - [3] The Decoder — 'Cloudflare says its new Clef model means humans no longer need to be in the loop for AI agents' (2026): https://the-decoder.com/cloudflare-says-its-new-clef-model-means-humans-no-longer-need-to-be-in-the-loop-for-ai-agents/ - [4] DataNorth — 'Cloudflare releases Clef and Clef Flash: open-weight decision models on Workers AI' (2026): https://datanorth.ai/news/cloudflare-releases-clef-and-clef-flash - [5] Slashdot — 'Cloudflare Tries to Outplay Jev With Open-Weight Clef Models' (2026): https://tech.slashdot.org/story/26/10/01/2110250/cloudflare-tries-to-outplay-jev-with-open-weight-clef-models --- _Canonical article: [https://brainwavepost.com/article/cloudflare-clef-decision-models-2026](https://brainwavepost.com/article/cloudflare-clef-decision-models-2026) — © BrainWavePost. Educational content; see the article page for full disclaimers._