Cloudflare releases two open-source Clef decision models on 1 October

Cloudflare released two decision models, Clef and Clef-flash, on 1 October 2026, hosted on Workers AI and open-sourced under an Apache 2.0 licence on Hugging Face. Decision models return bounded, structured answers with probabilities rather than free-form text. Cloudflare also debuted a reinforcement learning product that lets customers fine-tune Clef on their own data, alongside the model release.blog.cloudflare.com+1

The release matters because decision models offer deterministic, cheap, fast classification that can slot into an agent workflow where a decision is needed, unlike large language models that are largely non-deterministic. Clef adds a vision encoder for images and a 64k context window, versus Jev's text-only classification and 32k state-plus-question limit. Cloudflare's API is fully Jev compatible, so it can act as a drop-in replacement.techcrunch.com

What the release includes

Clef uses specially post-trained, frozen versions of Qwen3.8-27B and Qwen3.5-9B for Clef and Clef-flash respectively, with the Qwen backbone running a prefill-only pass during inference, then scoring choices in parallel. Both models answer three types of bounded questions: yes/no, multiple choice, and rankings. Jev's underlying architecture has not been disclosed by TypeSafe.

Cloudflare's Threat Intelligence team tested Clef on classifying website domains. Given a domain with Browser Run, Clef returned categories such as 95% chance of a fashion website, 85% ecommerce, and under 1% phishing, in 2.2 seconds to fetch, render and classify. Cloudflare's fastest general language model, gpt-oss-120b, took 4.7 seconds on the same workflow and returned only two classifications.

Cloudflare reported benchmark scores across the Jev Decision Index: on BFCL case-exact, Clef scored 98.47 and Clef-flash 98.76, against Jev's 95.75. On BANKING77 macro-F1, Clef scored 94.20 versus Jev's 79.74. On When2Call accuracy, Jev led at 80.97 against Clef's 72.37. On PhishNChips accuracy, DiffusionGemma Jev led at 85.35 against Clef's 79.60.

Where the benchmark numbers came from

On median latency, Clef recorded 209.3 milliseconds and Clef-flash 38.8 milliseconds, against Jev's 524.1 milliseconds and Laya's 5.8 milliseconds. Across 43 evaluation benchmarks, Cloudflare said its Clef models beat the decision models on latency except Laya, which trades off quality. Cloudflare also ran Typesafe's own eval suite, claiming Clef beat Jev in three of four areas, losing only on agent trace observability.

Clef costs $0.24 per million tokens, nearly six times Jev's $0.042 per million tokens. Cloudflare AI Platform group product manager Michelle Chen said that Clef-flash runs on any GPU with at least 41 GB of VRAM, while Clef requires 85 GB, assuming single concurrency and a 64k context window. Chen also confirmed the training datasets are not public.

Cloudflare self-reported its own scores against the benchmark, and they have yet to be reproduced for ranking on the official Decision Index. TypeSafe CEO and founder Diogo Almeida said he did not see real competition for his company emerging yet, calling the current batch more like machine learning people wanting to implement a cool architecture than teams making intelligence useful.

What remains unconfirmed

Amazon Web Services' Strands Labs released its own decision model, Strands Decider 2B, the same week, built on the torso of Qwen3.5-2B. Amazon distinguished engineer Marc Brooker said the model offers a workflow step structured to be more reliable through confidence scores and a closed domain of answers. Dozens of similar models have appeared since TypeSafe debuted Jev.

Cloudflare's reinforcement learning product lets customers fine-tune Clef on their own data for their use cases, released alongside the two models on 1 October 2026. The models are downloadable from Hugging Face under Apache 2.0 terms, the same licence as Qwen, for developers with hardware able to meet the 41 GB or 85 GB VRAM requirements.