Cloudflare Launches Fast AI Models for Automated Decisions

2026-10-03 companies

San Francisco, Saturday, 3 October 2026.
Cloudflare released Clef and Clef-flash, open-source AI decision models capable of making operational choices in just 39 milliseconds without needing human supervision or generating text.

Cloudflare Unveils Autonomous AI Decision Models

Cloudflare (NYSE: NET) announced the launch of its new open-source AI decision models, Clef and Clef-flash, designed to enable autonomous AI agents to make structured operational decisions without requiring human oversight or text generation [1]. The release occurred on October 1, 2026, marking a significant shift toward fully automated enterprise workflows and lower-latency edge AI infrastructure [3]. Built on the Apache 2.0-licensed Qwen framework, these models are hosted on Workers AI and aim to fill a niche between large language models and traditional classifiers by outputting probabilities rather than text [2]. This development allows agents to programmatically gather context, make decisions, and take actions on tasks, or defer to a human when needed [1].

Performance Benchmarks and Latency Analysis

In terms of speed, the Clef-flash model processes decisions in approximately 39 milliseconds, outperforming rival TypeSafe AI’s Jev model significantly [1]. Comparative benchmarks indicate the Jev model operates at over 524 milliseconds, meaning Clef-flash is approximately 13.436 times faster in median latency [1]. On the Decision Index, Clef achieved a score of 61.2, while Clef-flash scored 57.1, compared to the Jev model’s 57.9 [1]. Cloudflare claims the new models outperform competing solutions across 43 benchmarks, utilizing edge data center proximity to reduce network latency [1].

Technical Architecture and Capabilities

The Clef model is a 27B multimodal autonomous decision model based on the Qwen3.8-27B architecture, while Clef-flash utilizes Qwen3.5-9B for latency-critical tasks [4]. Unlike text-only competitors, Clef features a vision encoder and supports image processing, accepting inputs as text, JSON, images, or video [3]. The architecture utilizes a non-autoregressive, two-stage attention routing process that derives schema choices directly from internal backbone representations, bypassing token-by-token generation [2]. Technical specifications require torch 2.11 and transformers 5.10, with support for a 64,000-token context window [4].

Market Positioning and Availability

Cloudflare has made the model weights publicly available on Hugging Face under an Apache-2.0 license for external exploration and local deployment [2]. Pricing for hosted usage on Workers AI is set at $0.24 per million tokens, which is higher than Jev’s $0.042 per million tokens [5]. Cloudflare is also rolling out a reinforcement learning service for custom fine-tuning of Clef, initially supported by forward-deployed engineers [1]. While Cloudflare describes the models as open source, the training datasets remain private, and the company is seeking design partners for specific use cases [5].

Sources

AI agents Cloudflare