HaeB / Wikimedia Commons (CC BY-SA 4.0) The company's first internally trained open-weight models target fast, structured decisions for agentic workflows
Cloudflare has trained its own AI models for the first time. On October 1, 2026, the company introduced Clef and Clef-flash, two open-weight decision models now live on its Workers AI platform.
These models are built to make quick, structured calls, such as sorting a request into a category or deciding where a task should go next.
Two models, two speeds #
Clef is the heavyweight of the pair. It packs 27 billion parameters and is built on Qwen3.8-27B.
It also offers a 64k-token context window and accepts several input types: text, JSON, images and video.
Clef-flash is the lighter sibling, with 9 billion parameters and a foundation of Qwen3.5-9B. Cloudflare aims it at latency-critical paths, the parts of an application where waiting even a fraction of a second is too long.
Both models include a vision encoder that can process up to 4 images per request. Cloudflare says the models are optimized for outputs that are fast, accurate and deterministic.
The weights are fully open-sourced on Hugging Face under the Apache 2.0 license.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The numbers Cloudflare is leaning on #
Pricing is set at $0.24 per million input tokens for Clef and $0.09 per million input tokens for Clef-flash. Output tokens are not billed at all.
On accuracy, Cloudflare points to the BANKING77 classification benchmark. Clef scored a macro-F1 of 94.20, compared with 79.74 for Jev, a product from Typesafe.
On speed, Clef-flash posted a median latency of 38.8 ms. Jev came in at 524.1 ms.
Both models are compatible with the Jev-API. That means teams already built around Jev may be able to swap in Cloudflare’s models without rewriting their integrations.
Background: from network to model maker #
The launch landed during Cloudflare’s Birthday Week, the company’s annual stretch of product announcements. Workers AI has hosted models before, but Clef and Clef-flash are the first that Cloudflare trained in-house.
Alongside the models, Cloudflare is introducing a reinforcement learning fine-tuning platform. The company says it is set to launch with hands-on support for customers.
The models also run on Cloudflare’s edge GPU resources. The edge refers to servers located close to end users, which cuts the distance data has to travel and helps keep latency low.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our