Robots Atlas>ROBOTS ATLAS
Cloudflare Clef

Cloudflare Clef

27B · Family: Clef
Cloudflare's open-weight multimodal decision model (27B). Rather than text, it returns a probability distribution over allowed options from a state and a typed-question schema.
✓ Active✓ Public access⚖ Open weightsMultimodalSpecialized AI📁 Clef
Parameters
27B
parameters
Access:DownloadAPIDeployment:💻 Local☁ Cloud

Overview

Clef is an open-weight, multimodal decision model trained by Cloudflare. The main model (Cloudflare/clef) has 27B parameters and accepts text, JSON, images and video (Hugging Face task: image-text-to-text). The weights are released as BF16 safetensors.

Instead of generating free text, Clef takes a state and a schema of typed questions and produces decisions: it returns a logit for each allowed option and then a probability distribution via softmax.

The architecture is built on a Qwen3.8-27B backbone with a vision encoder, plus a lightweight shared schema head that reads the final hidden states and scores options jointly. The model was produced by post-training the Qwen/Qwen3.8-27B base model.

Clef is released under the Apache-2.0 license and its API is compatible with Jev and SystemOne. The model is part of the Clef family, alongside the smaller and faster clef-flash variant (9B) and the larger clef-omni (35B).

Classification
MultimodalSpecialized AI
Family: Clef
Access & deployment
DownloadAPI
LocalCloud
Weights: Open weights
Key parameters
🧩 Parameters: 27B
✓ Tools
📥 Input: text, structured data, image, video

Technical specification

Parameters
27B
parameters
License
Apache-2.0
Hardware requirements
BF16 weights (safetensors), ~27B parameters.
Features:✓ Tool use
Modalities
⬇ Input
textstructured_dataimagevideo
⬆ Output
structured_data

Capabilities and applications

Native model capabilities
Decision making
The ability to choose the best option from a defined set of possibilities based on input state, usually with a probability estimate for each option.
Category: reasoning
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Function Calling
Category: planning
Multimodal understanding
Category: multimodal
Image understanding
Analysing and interpreting the content of images.
Category: vision
Video understanding
The model's ability to analyse and interpret video content — recognising actions, motion, events and relationships between objects over time.
Category: video
Agentic capability
The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.
Category: planning

Benchmark results

6 benchmarks
BFCL (Berkeley Function-Calling Leaderboard)
accuracy
98.5%
📄 Decision Index 0.2.1 / Typesafe Evals
MMLU
accuracy
90.3%
📄 Decision Index 0.2.1 / Typesafe Evals
GPQA
accuracy · GPQA Diamond
48.0%
📄 Decision Index 0.2.1 / Typesafe Evals
ForecastBench
Brier score
13.9
📄 Decision Index 0.2.1 / Typesafe Evals
Lower is better.
safeagents-judge-bench
balanced acc 0.81, AUROC 0.99
📄 Decision Index 0.2.1 / Typesafe Evals
typed-decisions
accuracy
0.73
📄 Decision Index 0.2.1 / Typesafe Evals

Technical architecture