clef

Overview. Clef is a 27-billion-parameter decision model built by Cloudflare on a Qwen3.8-27B backbone paired with a custom joint schema head, released as open weights under Apache-2.0 on October 1, 2026. Instead of generating free-form text, Clef performs a single non-autoregressive pass over a defined set of options and returns calibrated, structured probabilities. It accepts text, JSON, and image inputs, with the open-weight repository also supporting video frames. Developers reach for it when an application needs a fast, schema-constrained decision rather than a conversational response.

Key strengths. Cloudflare trained Clef's routing head and low-rank adapters specifically for calibration, combining label-smoothed cross-entropy and Brier loss with a reinforcement-learning objective aimed at producing well-calibrated decisions rather than fluent prose. That focus is also its boundary: Clef has no free-form generative capability, making it unsuitable for code generation, writing, math, or agentic tool-loop tasks that depend on generated text. Its strength is narrow and deliberate — turning text, JSON, and image inputs into structured, probability-scored answers over a fixed option set.

When to use it. Use Clef when you need a calibrated classification or routing decision embedded directly in an application — for example scoring a fixed set of labeled outcomes from text, JSON, or an attached image — rather than a generated response. It's served through Workers AI, Cloudflare's edge Inference platform, making it a fit for latency-sensitive decision steps inside a larger pipeline. It is not the right choice for chat, content generation, or any task expecting free-form output.

Limitations. Clef produces no free-form text, streaming output, or reasoning traces, so it cannot substitute for a generative chat model. Cloudflare's own documentation specifies a 65,536-token context window for the hosted endpoint; the underlying backbone configuration declares a larger position capacity, but that has not been confirmed as a supported serving limit. No training-data cutoff date has been disclosed.

Sovereignty note. Clef is created and hosted by Cloudflare, a US-headquartered company; no creator-disclosed CLOUD Act exposure or training-on-input policy was found in the available documentation. Because Clef's weights are published openly under Apache-2.0, teams with data-residency requirements can self-host the model rather than relying solely on the hosted endpoint.