Clef and Clef-flash put Apache-2.0 weights, multimodal inputs, a compatible typed API, and local deployment behind models built to score bounded decisions instead of generate text.
Cloudflare's Clef family turns a relatively new model interface into something developers can download, self-host, and swap between implementations.
Clef and Clef-flash are decision models. Instead of generating free-form text, they receive state plus typed questions and return probabilities over allowed answers. A workflow can ask whether a request is urgent, which team should receive it, or how severe an event is, then act on the bounded result without parsing prose.
Cloudflare released both models on October 1. Their weights are available on Hugging Face under Apache 2.0: Clef is based on Qwen3.8-27B, while Clef-flash uses Qwen3.5-9B. Both preserve multimodal input and use the same typed decision interface.
That combination matters for the open ecosystem. The useful artifact is not only a pair of checkpoints. It is an interoperable model class with downloadable weights, a compatible API, and enough implementation detail to run outside Cloudflare's hosted service.
A conventional language model can be prompted to return JSON, but it still produces output autoregressively, token by token. Clef does not generate an intermediate textual answer and then convert it into a decision.
Cloudflare says the models derive schema choices directly from internal backbone representations. Each valid choice extracts relevant context through a two-stage attention-routing process, with field representations cross-attending to one another and back to the input before scoring. The result is a probability distribution over bounded options.
The public API exposes three question types: a yes/no-like noul decision, a choice over named alternatives, and a score over an ordered rubric. Multiple questions can be evaluated against the same state.
For agent systems, this separates two jobs often given to the same LLM. A generative model can still plan, write, or call tools, while a smaller decision model handles routing, escalation, or classification. Whether that separation improves a particular system depends on the workload, but it gives developers a concrete alternative to spending generative tokens on every bounded decision.
This is an important difference from model announcements that promise an open release later. The Clef model repository contains Apache-2.0 weights for a roughly 27B-parameter model, and Clef-flash provides the corresponding roughly 9B checkpoint under the same license.
Cloudflare froze the Qwen backbones and post-trained rank-256 low-rank adapters together with a specialized routing head. The training objective combines label-smoothed cross-entropy for valid schema outputs with Brier loss, which penalizes poorly calibrated probabilities. Cloudflare also describes synthetic training data that permutes prompts, field order, and schema structures.
The models keep a vision encoder, so state can include visual input rather than only text. Cloudflare's hosted implementation documents a 64K context window, compared with the 32K figure it cites for Typesafe's Jev.
Interoperability may be more consequential than either model size. Cloudflare made the interface compatible with Jev's decision API rather than introducing a separate request format. That means the ecosystem can start treating a decision model as a replaceable component instead of a single vendor endpoint.
There is already evidence of that happening outside Cloudflare. Relevan's open decision-model-eval implements Clef, Jev, Laya, CLM, and a structured-output Claude adapter behind one DecisionModel interface. Its Clef backend can point either to Workers AI or to a local server around the downloaded weights.
Cloudflare reports strong results across its evaluation set. In its published table, Clef or Clef-flash leads several classification and tool-oriented benchmarks, including BFCL, API-Bank, BANKING77, and CLINC150+OOS. It also reports median latency of 209.3 ms for Clef and 38.8 ms for Clef-flash, versus 524.1 ms for Jev in the same 43-benchmark test setup.
Those numbers should remain labeled as Cloudflare's measurements. They are useful for understanding the design target, not as independent proof of a universal performance advantage. Results vary by task: in Cloudflare's own table, Jev is ahead on When2Call and BRIGHT, while Clef-flash falls sharply behind the larger Clef on CLINC150+OOS.
A small independent test gives a useful sanity check. Relevan added local Clef results to its search-configuration evaluation on October 2. On ten index configurations, Clef produced nine completely correct configurations versus ten for Jev and a structured-output Claude setup. It got 34 of 35 enable decisions correct and all 15 encoder choices correct. Its Brier score was 0.015 in that run.
The latency result went the other way from Cloudflare's hosted benchmark: local Clef on an NVIDIA GB10 had a median call time of 0.72 seconds, while hosted Jev measured 0.16 seconds. The hardware and serving paths are different, and Relevan explicitly warns that latency will vary. Model-level benchmark latency does not directly predict an application's end-to-end timing.
The Relevan evaluation is narrow, and its labels are engineering hypotheses rather than established ground truth. It is useful because the code, cases, scoring rules, and local Clef setup are inspectable, not because 35 decisions settle the comparison.
The release itself is directly inspectable. The checkpoints exist, their Apache-2.0 licensing is explicit, the base-model lineage is recorded, and an independent repository documents a working local deployment path. Developers with sufficient GPU memory can reproduce the basic behavior without using Workers AI.
The broad performance claims require more caution. Cloudflare publishes detailed benchmark tables and methodology context, but they are still first-party results. The independent evidence available so far is much smaller and workload-specific. The larger Clef checkpoint is also heavy enough that local experimentation is not laptop-class by default; Relevan reports about 52 GB of downloaded model data and roughly 56 GB of GPU memory for its BF16 setup.
Open weights make it possible for independent evaluations, alternative runtimes, quantizations, fine-tunes, and API-compatible implementations to accumulate after release.
Clef therefore matters as infrastructure as much as a model launch. Typed decision models now have two Apache-licensed Cloudflare checkpoints, a hosted implementation, multimodal support, and an interface already shared with another provider. If that interface continues to spread, agent developers gain a component boundary where models can be compared and replaced without redesigning the workflow around each one.
