clef-flash

Overview. Clef-flash is an open-weight decision model from Cloudflare, post-trained from a Qwen3.5 9B backbone. Rather than generating free-form text, it returns calibrated probabilities over a fixed set of allowed options for a given schema, making it a purpose-built classifier rather than a general chat model. Developers who need a fast, schema-bound decision step — not a conversational response — are the intended audience.

Key strengths. The model pairs its backbone with a joint schema transformer head that constrains every answer to the options defined by the caller's schema, giving it reliable structured output and JSON-mode behavior by design. It accepts text, JSON, images and video frames as input, so it can make decisions that depend on visual context as well as text. Cloudflare describes training it with label-smoothed cross-entropy, Brier loss and reinforcement learning aimed at calibration, which points to probability outputs that are tuned to reflect real confidence rather than raw logits.

When to use it. Reach for Clef-flash when you need a simple, well-scoped decision — classification, routing, moderation-style option selection — made quickly at the edge, especially when the input includes an image alongside text or structured data. It is a good fit for simple, bounded-complexity workloads where a known set of outcomes is defined in advance. It is not the right choice for open-ended generation, multi-step reasoning, or conversational use cases, since it does not produce free-form text or visible reasoning traces.

Limitations. Clef-flash does not support tool calling, streaming, or extended thinking, and it has no published maximum for output tokens because it does not generate text token-by-token. Context window figures vary by source: Cloudflare's launch material advertises 64K tokens, but the current Workers AI endpoint documentation caps it at 24,576 tokens, so treat the smaller figure as the practical limit when deploying there. No training-data cutoff date has been published.