Clef: Cloudflare’s 27B Multimodal Model for Structured Decisions
Cloudflare’s Clef is a 27B multimodal decision model designed to return probabilities for predefined options instead of generating free-form text. That design changes how a system can represent a decision: text, JSON, images, and video can enter the model, while one forward pass produces the allowed-option probabilities for every question.
Cloudflare’s Clef is a 27B multimodal decision model designed to return probabilities for predefined options instead of generating free-form text. That design changes how a system can represent a decision: text, JSON, images, and video can enter the model, while one forward pass produces the allowed-option probabilities for every question.
Clef replaces open-ended generation with bounded outputs
Clef is post-trained from Qwen/Qwen3.8-27B and adds a joint schema transformer head to its 27B multimodal architecture. Its input can be state data represented as text or JSON, as well as images or video. Its output is not free-form text: for every question, Clef returns probabilities for each allowed option in a single forward pass. The model supports noul, choice, and score question types, giving the interface a defined schema before inference begins. That model-level choice leads directly to the way applications connect to it.
The interface is built around questions and allowed options
Clef’s output consists of one logit per allowed option for each question, and the output does not require output parsing. It is compatible with the Jev and SystemOne APIs, so developers working with those interfaces have a stated integration path rather than needing to treat the model as a general text generator. The model is therefore aimed at systems that can express their decision request through the supported question types and option schema. The next question is whether this constrained design performs well enough to justify using it.
Reported results show strong decision accuracy with sub-second latency
In the reported Decision Index 0.2.1 run, Clef achieved 98.5 percent BFCL case-exact accuracy and 90.3 percent MMLU accuracy. The same report gives a median latency of 209.3 milliseconds and a p95 latency of 238.6 milliseconds. These figures describe a model evaluated both on decision-oriented case exactness and on a broader knowledge benchmark, while the latency measurements provide a concrete reference for systems where response time matters. They also make the release practical to inspect rather than only conceptual.
An open release gives developers a concrete evaluation route
Clef is released under the Apache-2.0 license, and its model page provides the primary access point for examining the 27B checkpoint and its documented interfaces. The reported test environment used torch 2.11 and transformers 5.10.2 on one H200, which gives developers a specific software and hardware reference when reproducing the stated results; it is not a guarantee of identical performance in another environment. Clef-Flash is listed as a smaller variant. Taken together, the release offers a clear contrast with free-form generation: Clef accepts multimodal state data and returns schema-bound probabilities, so its practical value depends on whether your application can define its questions and allowed options precisely.
Specs
- Parameter count: 27B
- Backbone: Qwen/Qwen3.8-27B with vision encoder
- Input formats: Text, JSON, images, or video
- Output: One logit per allowed option for each question
- Maximum input length: 16,384 tokens by default
- Question types: noul, choice, and score
- License: Apache-2.0
- Smaller variant: Clef-Flash
Sources
- Cloudflare Clef model page — Hugging Face
- Decision Index 0.2.1 results — Cloudflare