Cloudflare Open-Sources Clef 9B and 27B Multimodal Decision Models With 38.8ms Edge Latency

Cloudflare has open-sourced Clef and Clef-Flash, a pair of 27-billion and 9-billion-parameter multimodal decision models engineered to sit directly in the network hot path and classify agent states, images, video, and JSON schemas in a single forward pass. Hosted across Cloudflare’s global edge GPU network and released with open weights, the smaller Clef-Flash 9B model achieves a median latency of 38.8 milliseconds—making real-time agent routing, bot detection, and content triage fast enough to execute inline on every HTTP request.

To understand what sets Cloudflare’s Clef apart from earlier decision models like TypeSafe AI’s Jev or AWS’s 2B text-only Strands Decider, look at two architectural upgrades: native vision multimodality and a 64,000-token context window. When an enterprise agent operates across web browsers, customer support portals, or security telemetry, the context it needs to evaluate is rarely a short text snippet. A support triage agent may need to inspect a user-uploaded screenshot of a billing error alongside a 40,000-token JSON session log, while a trust-and-safety agent may need to evaluate video frames alongside page DOM trees.

Whereas TypeSafe’s Jev is limited to text inputs and a 32,000-token context window, Cloudflare equipped both the 27B Clef model and the 9B Clef-Flash model with a native vision encoder and a 64,000-token context window (2x Jev’s capacity). Instead of generating text token-by-token, Clef ingests the multimodal state plus a schema of typed multiple-choice or boolean questions and outputs a calibrated probability distribution across every allowed option for all questions simultaneously in one forward pass.

On Cloudflare’s edge GPU benchmarks, the flagship 27B Clef model returns multi-question typed decisions in a median 209.3 milliseconds, while the latency-optimized 9B Clef-Flash model completes the same workflow in a median 38.8 milliseconds—more than five times faster than the 27B model and fast enough to run inside a Cloudflare Worker before an origin server even receives the request.

As reported by Renato Losio in InfoQ’s coverage of Cloudflare’s open-source Clef decision models, Cloudflare machine learning leads Michelle Chen, Alex Reneau, and Kevin Flansburg also announced a fine-tuning platform to adapt Clef checkpoints on historical labeled datasets for domain-specific bot classification, fraud scoring, and support routing.

Fast Facts: Cloudflare Clef & Clef-Flash Specifications

• Model Sizes Released: Clef (27 billion parameters) and Clef-Flash (9 billion parameters), both released as open-weight models.
• Multimodal Input Support: Processes text, structured JSON, images, and video frames via a built-in vision encoder, plus a schema of typed questions.
• Context Window: 64,000 tokens (double the 32,000-token window of TypeSafe’s Jev).
• Edge Benchmark Latency: 38.8 milliseconds median latency for Clef-Flash (9B) and 209.3 milliseconds median latency for Clef (27B).
• Single-Pass Execution: Returns calibrated probabilities for every allowed option across multiple typed questions in a single forward pass.

Comparing 2026 Agent Decision Models: Clef vs. Jev vs. Strands Decider

Why Single-Forward-Pass Multimodal Classification Changes Edge Architecture

When a traditional vision-language model (VLM) evaluates an image and answers five policy or routing questions, it must run an initial prefill pass over the image and text tokens and then run dozens of sequential autoregressive decode steps to spell out the JSON keys and values. Memory-bandwidth bottlenecks during that sequential decode phase are what push VLM latencies into the 800-millisecond to 3-second range.

Because Clef eliminates autoregressive text decoding and computes option probabilities for all typed questions in a single forward pass, developers get the visual and contextual understanding of a 9B or 27B multimodal model at the speed of a traditional classifier (38.8ms).

The 4-Step Enterprise Implementation Playbook

1. Move Agent Triage and Intent Routing to the Network Edge: Deploy Clef-Flash (9B) at the edge (via Cloudflare Workers AI or your ingress GPU tier) to classify user intent, inspect uploaded attachments, and route requests to the right backend agent in 38.8 milliseconds before waking up expensive origin compute.
2. Batch Multiple Policy and Routing Questions Into One Forward Pass: Take advantage of Clef’s typed schema input by passing all required checks—such as prompt-injection risk, PII presence, language detection, and department routing—in a single request rather than chaining sequential classifier calls.
3. Use the 64K Context Window for Full-Session Visual and Log Auditing: Feed complete multi-turn browser agent DOM snapshots and screenshots into the 27B Clef model (209.3ms) to verify whether a web agent is stuck in a UI loop or about to click an unauthorized button.
4. Fine-Tune Open Weights on Historical Support and Fraud Logs: Rather than relying on zero-shot prompt definitions alone, use your organization’s historical ticket-routing and abuse-moderation labels to fine-tune a private Clef-Flash checkpoint.