Cloudflare releases Clef, open-weight AI models that make yes-or-no decisions fast

Most AI chatbots answer by writing text, one word at a time. That's great for essays. It's slow and wasteful when an agent just needs a yes or a no.

A decision model skips the writing. Cloudflare explains it this way: Clef reads an input and a set of typed questions, then returns a probability for every allowed answer. The agent gets a structured decision it can act on right away, like "route this ticket" or "block this request."

Think of it like a multiple-choice exam versus an essay exam. The decision model just fills in the bubble.

Clef answers three kinds of questions, according to Cloudflare:

• Yes/no: the probability the answer is yes.
• Choice: pick one option from a list you define.
• Score: rate something against a scale you set.

You can ask up to 64 questions in a single request.

This category took off fast. The Register says Clef arrived two weeks after TypeSafe's Jev model "took the AI world by storm."

Startup Fortune explains the appeal with one example: a hackathon demo reportedly used Jev to monitor an AI agent's actions for $2.94, compared with $372 running the same check through a big frontier model.

Clef is also not alone this week. Startup Fortune reports that OpenAI introduced a Decisions API at its Dev Day on September 30, and Amazon released its own small decision model, Strands Decider 2B, on October 1.

Here's how the three compare, based on Startup Fortune's reporting:

Of the three, Cloudflare is the only one giving away the model weights at two sizes, Startup Fortune notes.

Startup Fortune adds some detail on Amazon's entry. Strands Decider 2B came out of Amazon's Strands Labs and is built on a Qwen3.5 base. It reportedly started as an internal prototype from AWS distinguished engineer Marc Brooker, and scores around 72% on the public JevBench v19 test set.

Its takeaway for developers: the real cost isn't only the price per answer, it's lock-in. A team that builds on a closed API depends on one company's uptime and pricing. Open weights let you run the model wherever you want.

• Clef: a 27 billion parameter model aimed at the most accurate answers.
• Clef-flash: a 9 billion parameter model built to answer in a few tens of milliseconds.

The Register reports both are built on frozen, post-trained versions of Alibaba's Qwen models. Instead of generating words, they read the question once and score the possible answers directly. Startup Fortune calls this a "non-autoregressive, prefill-only" approach.

That's a contrast with Jev, whose design TypeSafe has kept secret, The Register notes.

Both Clef models support a 64k token context window. Jev also handles up to 64k tokens per request, The Register adds, but its state plus its longest single question is capped at 32k. Clef can also look at images: Cloudflare says it has a vision encoder and accepts up to four images per request. The Register adds that it handles video too, something Jev doesn't.

By Cloudflare's own numbers, yes. Across 43 benchmark runs, Cloudflare says Clef was 2.5 times faster than Jev at the median, and Clef-flash 13 times faster.

On accuracy, Cloudflare says a Clef model scored highest on 7 of 10 decision benchmarks. On TypeSafe's own workflow tests, Clef beat Jev in 3 of 4 areas: invoice processing, customer service and security incidents.

The catch: The Register points out that Cloudflare self-reported those scores. They hadn't yet been reproduced for the official Decision Index ranking. The one area Clef lost, The Register adds, was agent trace observability. The Register also notes a trade-off in Cloudflare's own ranking: Clef is a little slower than some other open decision models but more accurate, while Clef-flash matches most of them on accuracy and is far faster.

One real-world test from Cloudflare: paired with its Browser Run tool, Clef fetched, rendered and classified a website in 2.2 seconds, versus 4.7 seconds for the gpt-oss-120b model in the same workflow.

You have two options.

Hosted: Clef runs on Cloudflare's Workers AI. Cloudflare says that adds speed, because requests run on GPUs across its network, close to users. The Register reports the price: $0.24 per million tokens, nearly six times Jev's $0.042.

Self-hosted: the weights are on Hugging Face under the Apache 2.0 license, so you can run Clef on your own hardware. Cloudflare's Michelle Chen told The Register what you'll need:

Chen said that assumes one request at a time and a 64k context window. She also confirmed to The Register that the training datasets aren't public, even though the announcement called the models open source.

Switching should be easy for Jev users. Cloudflare says Clef follows the same API, so you change the endpoint and model name.

Developers can call Clef through the Workers AI binding in code or a REST API, and route it through Cloudflare's AI Gateway. Cloudflare's suggested pattern is to put Clef directly in an agent's request path for the quick decision, then hand off to a larger language model on Workers AI when the agent needs to actually do something.

Cloudflare lists several uses:

• Support triage: decide if a ticket is urgent and which team owns it, then route it without a human.
• Trust and safety: score user posts against your own policy rules.
• Agent guardrails: let an agent ask "should I take this action?" in tens of milliseconds before calling a tool.
• Threat intelligence: classify websites by category.
• Visual checks: classify images, not just text.

In real life A customer writes "checkout has been failing for every customer for the last hour." Clef instantly marks it urgent and sends it to the technical team, before any human reads it.

Cloudflare is also launching a reinforcement learning fine-tuning service so customers can tune Clef for their own workloads. It's asking interested teams to sign up as design partners. Cloudflare had a busy week: it also made its Basin data platform generally available.

• If you build AI agents: you can try Clef on Workers AI today, or download it and run it yourself.
• If you already use Jev: Clef is a drop-in alternative, but compare prices, since hosted Clef costs more per token.
• If you care about lock-in: open weights mean you can keep running the model even if you leave Cloudflare's hosting.
• If you're just curious: these models are why more apps may soon make small decisions for you, quickly and quietly.