Cloudflare Open Sources Decision Models for AI Agents

Cloudflare announced Clef during its "Birthday Week," a set of open-weight AI models positioned to choose among predefined options rather than generate text. Alongside the models, it also released a platform for adapting them to specific decision tasks. This positioning itself sets Clef apart from mainstream generative models: it doesn't write answers, it selects them.

In concrete form, Clef is a 27B multimodal model whose input is a state and a schema of typed questions, and whose output is a decision. It can handle text, JSON, images, or video, and in a single forward pass returns the probability of each allowed option for each question, requiring neither free-text generation nor output parsing. Cloudflare released both 9B and 27B parameter versions.

This design corresponds to a specific step in agent workflows: classifying input into typed results with probabilities, so the agent can decide the next step—for example, routing a support request, escalating it, or handing the decision to a human. It should be noted that this "classify-and-route" use-case description comes from the report's explanation of the decision-model category, not from a capability claim unique to Clef.

Cloudflare also released Clef-Flash, a smaller 9B multimodal model aimed at latency-sensitive decisions. In Cloudflare's benchmark, Clef-Flash has a median latency of 38.8 milliseconds, while the 27B Clef model is 209.3 milliseconds. The roughly fivefold difference points to the same trade-off: parameter scale buys speed.

These numbers must come with qualifiers. They come from Cloudflare's own benchmark and are a source claim, with no third-party independent verification at present. That is, 38.8 milliseconds can be cited as "Cloudflare's claimed median latency," but not as a neutral measurement result.

The Cloudflare team also says that because the models are hosted on Cloudflare infrastructure, edge GPUs can bring low network latency and faster decisions, allowing Clef to be placed in an agent's hot path, paired with some LLM on Workers AI to execute actions. This edge advantage is likewise a source claim, and its performance in actual deployments has not been independently measured.

The Cloudflare team uses two dimensions to explain how Clef differs from other decision models. First, Clef has a vision encoder and can receive images and classify visual content, whereas Jev currently only does text classification. Second, Clef has a 64k context window versus Jev's 32k, meaning users can pack in more input states to classify. Both points are statements by the Cloudflare team and are source claims.

At the same time, Cloudflare's API is compatible with Typesafe AI's Jev System One model. This compatibility is a confirmed fact, but it also draws the boundary of ecosystem interoperability: Clef's integration path is constrained by compatibility with the Jev interface rather than establishing a separate one.

Putting these two sets of information together, Clef's differentiation narrative is "more multimodal and longer context than Jev," while its adoption path is "compatible with Jev." The former is Cloudflare's self-statement; the latter is an established fact at the interface level.

Cloudflare announced a fine-tuning service that allows customers to adapt Clef to their own workloads using their own data. Initially it is supported by Cloudflare engineers, with a self-serve platform planned for a later release.

The key constraint is timing: Cloudflare has not yet announced a specific date for the self-serve fine-tuning platform. This means teams that want to customize Clef at this stage effectively depend on intervention by Cloudflare engineers, limiting scalability.

Cloudflare plans to use its historical labeled data to fine-tune Clef for specific use cases such as support ticket triage and bot classification, in order to improve accuracy and speed. This plan is a source claim, and no corresponding results have been disclosed yet.

Cloudflare has made the Clef models available through Workers AI and provides downloadable weights on Hugging Face, inviting developers to try them and give feedback. This distribution arrangement is a confirmed fact.

The two channels correspond to two types of users: Workers AI is for developers who want to call the models directly within Cloudflare's runtime, while the Hugging Face weights are for developers who want to download, inspect, or even deploy them themselves. Open weights plus a hosted entry point form the complete shape of this release at the channel level.

In the related discussion on Hacker News, the community debated decision models, latency, open weights, and whether this is really a new model category. User Jacek Złydach said this is not a "new paradigm" but low-hanging fruit that has existed for years, and that Typesafe is simply the first to stop and pick it up and market it heavily. This is a community view and a source claim.

The benchmark issue was also called out directly. User SebastianSosa warned that public benchmarks are easy to game, and said that if he were Typesafe, he would also publish public benchmarks to distract attention, letting capable people overfit to the benchmarks rather than build something genuinely useful. This is likewise a community view and does not represent something proven.

There is one more doubt from Reddit worth noting separately: user bugra_sa said that rather than raw accuracy scores, he cares more about whether Clef knows when to "punt," and suggested testing it in scenarios where the cost of a false positive far exceeds that of a false negative, then changing the data to see when its confidence collapses. This points to an unanswered question: whether Clef can correctly abstain when the cost of a false positive far exceeds that of a false negative is currently unclear.

Separating confirmed facts from source claims, Clef's outline is clear: it did open-source both 9B and 27B versions, it is indeed positioned as "choosing rather than generating," it is indeed distributed through Workers AI and Hugging Face, and it does offer a fine-tuning service initially supported by engineers with no date for the self-serve platform. These are established facts at the release level.

Its performance narrative, however—38.8 milliseconds versus 209.3 milliseconds, 64k versus 32k, the vision encoder, the low latency from edge GPUs—all comes from Cloudflare itself or its team's statements, lacking third-party independent verification. The community's disagreement over "whether it is a new paradigm" and "whether public benchmarks are trustworthy" will not be resolved in the short term by Cloudflare's own numbers.

A more pragmatic judgment, therefore, is this: Clef offers an engineering option for pulling decisions out of generative models, and its interface is compatible with the Jev ecosystem, lowering the cost of trial and error; but whether it is truly faster, more accurate, and better at knowing when to abstain requires users to verify with their own data, rather than citing the publisher's benchmarks.