Cloudflare’s Clef Scores Agent Decisions With Vision and a 64k Window

On October 1, 2026, Cloudflare released Clef and Clef-flash—the first models trained by its Workers AI team—as hosted endpoints and Apache-2.0 open weights, positioning decision models as a first-class layer beside general LLMs for agent workflows. Where a frontier chat model spends tokens drafting free-form answers, a decision model takes a state plus a schema of typed questions and returns a probability for every allowed option in one forward pass. Cloudflare’s launch blog and Hugging Face cards frame Clef as Jev/SystemOne API-compatible, so teams already wiring bounded classifiers into routing, triage, or tool gates can swap the endpoint without rewriting clients.

Decision models sit in a different niche from reasoners. They sacrifice open-ended generation for speed, schema discipline, and calibrated scores that application code can branch on. Cloudflare’s Threat Intelligence vignette is concrete: with Browser Run, Clef fetched, rendered, and classified a domain in about 2.2 seconds—returning multi-label probabilities such as fashion, ecommerce, and phishing—while the company’s fastest general LLM in the same workflow, gpt-oss-120b, took about 4.7 seconds and returned only two classifications. That is the product pitch in one comparison: put a small, typed decision step in the hot path, then hand off to a larger model only when language generation is actually required.

Clef differentiates on modality and context. The larger model keeps the Qwen3.8-27B backbone and its vision encoder, so state can include text, JSON, images, or video; Clef-flash is the latency-oriented sibling post-trained from Qwen3.5-9B. Cloudflare cites a 64k context window versus Jev’s 32k, and publishes Decision Index 0.2.1 numbers that put Clef at or near the top of several routing and classification tasks—BFCL case-exact about 98.5%, BANKING77 macro-F1 about 94.2%, CLINC150+OOS about 97.4%—while Clef-flash’s median latency is listed at 38.8 ms against Clef’s 209.3 ms and Jev’s 524.1 ms in the same table. On Typesafe’s workflow suite, Cloudflare reports Clef beating Jev on invoice processing, customer service, and security incidents, with Jev still ahead on agent-trace observability. Those figures are company-run; the live Decision Index demo is the place to re-check as weights and evals move.

Under the hood, inference is deliberately non-autoregressive for the decision step. After a Qwen prefill, Clef scores valid schema choices in parallel from internal representations rather than emitting intermediate tokens. Cloudflare describes a joint schema head with two-stage attention routing so each option gathers evidence from the state, fields cross-attend, and scoring stays schema-bound. Post-training freezes the backbone, trains a rank-256 LoRA plus the routing head with label-smoothed cross-entropy and a Brier term for calibration, then applies Reinforcement Learning for Calibrated Decisions (RLCD) as a secondary objective that gives partial credit to adjacent ordinal choices. The Hugging Face Cloudflare/clef card documents systemone() for Jev-shaped request bodies, plus images / videos fields for multimodal records, and states Apache-2.0 licensing following the Qwen base.

Hosting and fine-tuning sit on Cloudflare’s agent-cloud stack. Both models are available as @cf/cloudflare/clef and @cf/cloudflare/clef-flash on Workers AI, with an enterprise guarantee that Cloudflare does not read, store, or train on request traffic unless the customer opts into fine-tuning. The same blog debuts a reinforcement-learning product path: AI Gateway to capture traffic as datasets, Containers as RL sandboxes, a Trainer to update weights, and Bring-Your-Own-Model redeploy on Workers AI—starting with forward-deployed engineers before self-serve. Internal use cases called out include Trust & Safety triage, support routing, and bot good/bad classification, where years of labeled decisions can specialize Clef beyond the general checkpoint.

Builders should still treat decision models as specialist tools. They are a poor fit for coding, summarization, or multi-step reasoning that needs free text; Cloudflare is explicit that Clef returns only structured probabilities. Accuracy varies by benchmark—Jev still leads some hard reasoning and When2Call rows in Cloudflare’s own table—so hybrid designs that reserve Clef for cheap gates and a larger LLM for hard turns remain the practical pattern. Weights and the Workers AI endpoints are available today; the open cards and changelog post are the durable primaries for license, sizes, and API shape.

For Writeble’s opensource desk, Clef is the clearest October 1 signal that multimodal, Jev-compatible decision weights are now a Cloudflare-native commodity rather than a single-vendor API niche.