Clef

Clef is a System One model trained by Cloudflare. You send a state and a set of typed questions, and it returns a probability for every allowed option of every question, with no generated text. It comes in two sizes. Clef is 27B parameters and Clef-flash is 9B. Both run on Cloudflare Workers AI as @cf/cloudflare/clef and @cf/cloudflare/clef-flash. Cloudflare calls the request format “fully Jev-API compatible”. It uses the same state and questions fields and the same Noul, Choice and Score question types as Jev, so the request body moves between the two with the model field changed.

Two differences are Cloudflare’s own list. Clef reads images, where Jev reads text only. The docs add an optional images array of up to four images, which they call a Clef extension to the System One API. Its context window is 65,536 tokens, against 32,000 for Jev per Cloudflare’s post. The docs also say the state can be text, JSON, images or video.

Cloudflare says each model keeps a frozen Qwen backbone, Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash. A routing head and rank-256 adapters are trained on top. Inference runs one forward pass over the input, then scores every valid answer in parallel, so no text is generated token by token. Training used label-smoothed cross-entropy with a Brier loss for calibration, synthetic data, and a second stage Cloudflare calls Reinforcement Learning for Calibrated Decisions. It builds on Cloudflare’s earlier DiffusionGemma experiment, which drew on Matt Mastracci’s djev. See how Jev is built for the comparison.

Every figure here is vendor-run: Cloudflare ran the evals and published the tables. On 10 tasks it picked from the Decision Index, Clef scores higher than Jev on eight and lower on two: When2Call, 72.37 against 80.97, and BRIGHT, 45.91 against 47.52. Clef-flash scores 66.77 on CLINC150+OOS against Jev’s 89.27. On TypeSafe’s workflow evals Clef beats Jev on invoice processing (64.7 against 61.8), customer service (76.3 against 76.0) and security incidents (62.9 against 61.7), and trails it on agent trace observability (68.5 against 71.6). Cloudflare also says Clef is currently the leader when evaluated against the Jev Decision Index. We have not seen an outside run.

Short typed calls where you want a probability to act on: routing a support request, picking a team, scoring severity, or sending a call to a human when confidence is low. Cloudflare’s example classifies a website’s category from a fetched page. Because it reads images, it can also classify a screenshot or photo.

It writes no text. Clef-flash scores well below Clef on some tasks, such as CLINC150+OOS (66.77 against 97.43, Cloudflare’s numbers). The latency figures are Cloudflare’s own and do not say how they were measured.

Both models are live on Workers AI, billed per input token. The weights are on Hugging Face under Apache 2.0, and Cloudflare also announced a fine-tuning service, hands-on at first and self-serve later. Cloudflare says it does not read, store or train on requests or responses unless you opt into fine-tuning.

Question types Choice Score Noul Max Choice optionsNot documented Score levelsNot documented Questions per call64 Total context65,536 tokens State budgetNot documented Rate limitNot published on the Workers AI model pages as of 2026-10-02. EndpointPOST https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/@cf/cloudflare/clef, or env.AI.run("@cf/cloudflare/clef", {...}) from a Worker SDKs Workers AI lists a 65,536-token context window for both models and says long text state is truncated to fit the token limit. A request takes 1 to 64 questions. The optional images array holds up to 4 PNG, JPEG or WebP images, each up to 4 MiB and 16 megapixels, 8 MiB in total, inside a 13 MiB request body. Remote image URLs are not accepted. The docs publish no cap on options per Choice or levels per Score.

@cf/cloudflare/clef, 1 Oct 2026, The 27B model, built on a frozen Qwen3.8-27B backbone. About 27.4B parameters per Hugging Face. Apache 2.0 weights as Cloudflare/clef. Release notes @cf/cloudflare/clef-flash, 1 Oct 2026, The 9B model for latency-sensitive calls, built on a frozen Qwen3.5-9B backbone. About 9.4B parameters per Hugging Face. Apache 2.0 weights as Cloudflare/clef-flash. Release notes

What people use Clef for, one page per pattern. Workflow controlStarter

Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket. Choice Score Noul Workflow controlIntermediate

One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue. Choice Score Workflow controlIntermediate