Cloudflare releases Clef decision models for AI agents

On October 1, 2026, Cloudflare released Clef and Clef-flash, its first in-house trained AI models, designed to make fast, deterministic routing decisions for AI agents. Built on Qwen backbones, the open-weight models run on Workers AI and compete directly with TypeSafe AI's Jev. Simultaneously, AWS Strands Labs launched Strands Decider 2B, a 1.9-billion-parameter open-source model that removes text generation from agent loops. These simultaneous releases signal a hardening infrastructure tier of 'System One' models optimized for speed and structured choices.

Cloudflare Clef specifications

• ▪Clef supports a 64k-token context window and can process text, images, and video, whereas TypeSafe AI's Jev is limited to text.
• ▪On October 1, 2026, Cloudflare released Clef and Clef-flash, two open-weight decision models, on Hugging Face under the Apache 2.0 license, though its training datasets remain private.
• ▪Clef is built on Qwen3.8-27B with 27 billion parameters, while Clef-flash is built on Qwen3.5-9B with 9 billion parameters.

Clef performance and compatibility with Jev

• ▪Cloudflare's Clef scored a macro-F1 of 94.20 on the BANKING77 classification benchmark, compared to 79.74 for TypeSafe AI's Jev.
• ▪Cloudflare reported a median latency of 38.8 milliseconds for Clef-flash and 209 milliseconds for Clef, compared to 524.1 milliseconds for TypeSafe AI's Jev.
• ▪Clef and Clef-flash feature API compatibility with Jev, allowing them to serve as drop-in replacements for teams already integrated with TypeSafe AI's API.

Cloudflare customization and infrastructure

• ▪To run custom models, Cloudflare uses technology from Replicate, which Cloudflare acquired in late 2025.
• ▪Cloudflare is introducing a reinforcement learning fine-tuning platform and service to let customers customize Clef for their own tasks.

AWS Strands Decider 2B specifications

• ▪On an NVIDIA RTX 3090, Strands Decider 2B answers questions in a median of 115 milliseconds, with a 95th-percentile latency of 299 milliseconds.
• ▪On October 1, 2026, AWS Strands Labs published Strands Decider 2B, a 1.9-billion-parameter model, on Hugging Face under an Apache 2.0 license with its training recipe, data list, and evaluation harness fully public.
• ▪Strands Decider 2B was fine-tuned using LoRA at rank 16, allowing the full training recipe to run in 11 hours on a single NVIDIA RTX 3090 GPU.
• ▪Strands Decider 2B replaces the traditional language-modeling head of Qwen3.5-2B-Base with a pointer head to compare hidden-state representations.

Overview of decision models

• ▪Decision models like Clef and Strands Decider 2B support three structured question types: yes/no (noul), multiple choice, and rankings (score).
• ▪Decision models solve agent pipeline bottlenecks by picking from bounded option sets without generating text, avoiding slow and expensive frontier LLM APIs.

Debatable claims

• ▪AI models built on Chinese base weights pose an unacceptable security risk to Western enterprises
• ▪Open-weight AI models are superior to proprietary closed systems for enterprise deployment
• ▪AI agents can safely make autonomous decisions without human-in-the-loop oversight