Cloudflare Clef Decision Model Family Launches: 38.8ms Ultra-Low Latency, Apache 2.0 Open-Weight, Reshaping AI Agent Classification Infrastructure

Key Highlights

In October 2026, Cloudflare officially launched the Clef decision model family, a suite of high-speed structured classification models designed for AI agent "hot-path" workflows. Unlike traditional large language models (LLMs) that generate free-form text, Clef models accept input states and directly return structured answers with probability distributions—yes/no, multiple-choice, or numeric scores—at speeds over 10x faster than standard LLMs.

The Clef family is built on the Qwen3 architecture, released as Apache 2.0 open-weight models available for self-hosting on Hugging Face, and natively deployed on Cloudflare's Workers AI edge network.

---

Three Models in Detail

Clef-flash: The Ultra-Speed Classification Engine

• Parameter Scale: 9B (9 billion parameters)
• Median Latency: 38.8 milliseconds
• Context Window: 64K tokens
• Primary Use: High-frequency text and JSON classification tasks
• Architecture: Uses a "prefill-only" approach, scoring all options in parallel without token-by-token generation

Clef-flash is the fastest of the three models, designed for scenarios requiring millisecond-level responses, such as real-time content moderation, user request routing, and initial security incident classification.

Clef: The Precision Reasoning Model

• Parameter Scale: 27B (27 billion parameters)
• Median Latency: 209.3 milliseconds
• Context Window: 64K tokens
• Primary Use: Complex reasoning, edge-case classification, and out-of-scope query filtering
• Architecture: Larger parameter scale provides higher classification precision for scenarios requiring nuanced judgment

Clef-omni: The Multimodal Decision Model

• Parameter Scale: 30B (30 billion parameters, 3B active)
• Release Date: October 9, 2026
• Supported Inputs: Text, JSON, images, audio (WAV/MP3), video (MP4/WebM)
• Architecture: Processes multimodal inputs in a single forward pass without multiple API calls

Clef-omni is the most groundbreaking of the three, combining multimodal understanding with decision output, enabling AI agents to make classification decisions based on combined visual, audio, and textual information.

---

Technical Architecture: Why 10x Faster Than LLMs?

Traditional LLM inference uses "autoregressive generation"—generating one token at a time until the complete response is finished. This approach is highly effective for long-form text output but represents enormous waste for classification tasks.

The Clef family uses a "prefill-only" architecture:

1. The model receives the input state (text, images, etc.) and a predefined list of options
2. In a single forward pass, it computes probability scores for each option in parallel
3. It directly returns structured results: the highest-probability option, probability distribution across all options, and a confidence score

This architecture completely eliminates the overhead of token-by-token generation, reducing latency from seconds to milliseconds.

---

API Design and Developer Experience

Clef's API design is clean and intuitive. Developers define a question schema, and the model returns structured results.

Three Output Formats:

1. Predicate: Returns a probability value from 0 to 1 for yes/no conditions

   • Example: "Does this image show damage?" → 0.87 (87% probability yes)

2. Choice: Selects one option from a fixed list, returning the selected value and probability distribution

   • Example: "Which category does this support ticket belong to?" → "Technical Issue" (confidence: 0.92)

3. Score: Returns a probability-weighted mean across ordered levels

   • Example: "What is the severity of this security incident?" → 3.7/5 (high severity)

The API is drop-in compatible with the "Jev System One" model interface, enabling seamless migration for developers already using that interface.

---

Open-Source Strategy: The Strategic Significance of Apache 2.0

Cloudflare's decision to release Clef model weights under Apache 2.0 license on Hugging Face carries profound strategic implications:

• Self-Hosting Flexibility: Enterprises can deploy on their own GPU infrastructure with complete data privacy control
• Community Ecosystem Building: Open-source licensing attracts developer community contributions and improvements
• Competitive Differentiation: Contrasts with closed-source competitors like OpenAI Decisions API, attracting enterprises prioritizing data sovereignty
• Fine-Tuning Roadmap: RL fine-tuning currently available through a design-partner program, with self-serve fine-tuning planned for the future

---

Benchmarks and Competitive Comparison

Cloudflare reports that the Clef family outperforms competitors across multiple classification tasks:

• Invoice Processing: Clef excels in both accuracy and speed
• Customer Service Classification: Clef-flash performs best in high-frequency scenarios
• Security Incident Classification: Clef demonstrates higher precision in complex edge cases

Notably, the Clef family directly competes with Microsoft Decision-1 (another decision model released around the same time). Microsoft Decision-1 is based on Qwen3.5-9B and priced at $0.042 per million input tokens; Clef is available through Cloudflare Workers AI with a more competitive pricing strategy and offers open-source self-hosting options.

---

Use Cases: The "Decision Brain" for AI Agents

The Clef family is best suited as the "decision layer" in AI agent workflows:

1. Content Moderation: Real-time determination of whether user-generated content violates policies, with millisecond-level response
2. Intelligent Routing: Directing user requests to the most appropriate processing flow or human agent
3. Agent Action Control: In multi-step workflows, deciding whether an agent should "continue," "retry," or "escalate"
4. Security Monitoring: Real-time classification of network traffic and API call risk levels
5. E-commerce Recommendations: Rapidly classifying purchase intent based on user behavior signals

---

Opportunities for Asia-Pacific Developers

For Asia-Pacific developers and enterprises, the Clef family launch presents important opportunities:

• Edge Deployment Advantage: Cloudflare has an extensive edge node network across Asia-Pacific, enabling Clef models to run closer to users for even lower latency
• Data Sovereignty: Apache 2.0 open-source licensing allows Asia-Pacific enterprises to self-host on local infrastructure, meeting each country's data localization regulations
• Cost Efficiency: Compared to using large LLMs for classification tasks, the Clef family can dramatically reduce AI inference costs—particularly important for cost-sensitive Asia-Pacific markets

---

Conclusion

The launch of Cloudflare's Clef family marks a new phase of "specialized division of labor" in AI infrastructure. Generative LLMs handle creativity and reasoning, while decision models handle high-frequency, structured classification tasks—each playing its role to form an efficient AI agent ecosystem.

For developers building AI agent applications, the Clef family offers a decision infrastructure choice that combines speed, precision, and openness. As AI agent deployments continue to scale in enterprises, efficient decision models will become indispensable foundational components.