Cloudflare opens Clef decision models as Jev clones multiply

The announcement, on the company blog, lands the same day Amazon pushed its own Jev-style model and hours after TypeSafe AI's Jev was described by InfoQ as a decision-only system that returns typed probabilities instead of text. Cloudflare says Clef currently leads the Jev Decision Index, the benchmark site it links to from the post.

Decision models are a narrower bet than the general purpose assistants that dominate headlines. They take a set of candidate answers, or "questions" with defined output types, and return probabilities rather than prose. Cloudflare's own framing is blunt about where the value sits: the model classifies, your code routes.

The company gave one worked example from its own operations. Feeding a domain through Clef with Browser Run, it classified a site at 95% fashion, 85% ecommerce and under 1% phishing. Cloudflare says the fetch, render and classify loop took 2.2 seconds for Clef, against 4.7 seconds for its fastest general LLM, gpt-oss-120b, which returned only two classifications in the same workflow. That is a latency claim, not an accuracy one, and it comes from Cloudflare's own Threat Intelligence team rather than an independent test.

Open weights, and a benchmark fight

Both models are hosted on Workers AI and released under Apache 2.0, which permits commercial use. That is a different posture from the closed APIs around most frontier assistants, and it puts Cloudflare in the same corner as the open-weight crowd it is competing with on cost and latency. The competitive field is getting crowded fast.

TechCrunch reported on 30 September that OpenAI used its Dev Day to show a Decisions API, described by CEO Sam Altman as a way to give the company's Luna model a predefined set of options to choose between. Altman said focusing the model on that choice keeps it fast while retaining image understanding, language coverage and safety protections.

TypeSafe AI CEO Diogo Almeida, a former OpenAI engineer, responded on X with a joke about the clone wars, TechCrunch reported, adding that OpenAI's move could be a sign that building in a System One compatible way is the future. Almeida told TechCrunch that his company's moat is synthetic data for statistically useful outputs, and that "fast and cheap is very easy" while intelligence is the hard part.

Not everyone agrees Jev is ahead

Independent comparisons are already muddying the picture. A GitHub write-up published on 25 September compared an unfinetuned Qwen3.5-9B against Jev 1.13.0 and reported comparable accuracy with lower average calibration error on several benchmarks, without decision-specific training. That project reports Jev leading on JevBench (86.1% against 80.5%), Typed Decisions (73.9% against 62.1%) and When2Call (73.3% against 61.8%), while Qwen3.5-9B led on phishing (73.9% against 62.5%) and MetaTool (83.4% against 77.1%). WebShop was near identical at 24.8% and 24.6%.

The same repository reports latency in Jev's favour on most tasks despite running Qwen on a single NVIDIA H100 80GB. That is the awkward part of the decision-model pitch: if a stock language model can be queried for next-token probabilities and land in the same range, the case for a separate model class rests on speed, calibration and cost per decision, not raw capability.

Another open project, Contrastive Language Models, claims a different trade-off. Its repository, last updated on 25 September, describes CLM-8B as pre-trained on 60M Nemotron Q&A pairs, mid-trained on 30M synthetic hard negatives and post-trained on 1M agentic trajectories, performing on par with Jev across computer-use, gaming and tool-calling with up to 9x lower latency, and reaching 87.6% on Terminal-Bench 2.1 and 81.6% on DeepSWE as a verifier after fine-tuning.

The safety angle

Cheap decision models have an obvious second life in oversight. TechCrunch reported that one of OpenAI's new security measures after a run of agent misbehaviour uses a separate model to watch for bad actions at significant compute cost, and that a long-time cybersecurity professional, Shapor Naghibzadeh, built a hackathon demo using Jev to check each agentic action against its assigned task.

That use case sits awkwardly next to the rest of the week's news. Ars Technica reported on 30 September that the nonprofit Legal Advocates for Safe Science & Technology sued OpenAI over the July 2026 hack of Hugging Face, arguing California law leaves no defence that an AI autonomously caused the harm. OpenAI called the suit completely without merit.

Cloudflare's own framing avoids the safety debate. Its post describes decision models as a way for agents to gather context, decide and act, or defer to a human when needed. Whether that deferral happens is a design choice in the customer's code, not a property of the model.

The company also said it is open-sourcing the two models on Hugging Face for local use, and debuting a reinforcement learning product that lets customers fine-tune Clef for their own categories. Both are available now.