Cloudflare's open-weight Clef models score every allowed option in a single pass

Clef takes a state and a schema of typed questions as input. The state can be text, JSON, images or video ^[2]. For each question, the model returns a probability for every allowed option in one forward pass ^[2]. It writes no free-form text, so nothing needs parsing ^[2]. The agent reads the scores and acts on them: it routes the support request, escalates it, or defers to a human ^[3].

For a support queue with a fixed set of destinations, I think this is the right design. The outcomes are a closed set, known before the call, and the schema writes them down ^[2]. With free-text generation, the agent has to parse an answer before it can act on it ^[2]. Clef returns scores against the schema, so the parse step goes away ^[2].

Cloudflare did not invent the format. Clef's API is compatible with Typesafe AI's Jev System One ^[4]. Matching a competitor's API is a polite way to concede who wrote the format. On Hacker News, Jacek Złydach wrote that "it's a low-hanging fruit that's been lying around for years; Typesafe were the first to bother to stop and pick it up" ^[11]. Cloudflare's engineers name two differences. Clef has a vision encoder, while Jev "only does text classification today," and Clef's context window is 64k tokens against Jev's 32k ^[10]. That gives it twice the room for input state ^[14].

Speed is the argument for the smaller model. Clef-Flash, at 9B parameters, is built for latency-sensitive decisions and posted a median of 38.8 ms in Cloudflare's benchmarks, against 209.3 ms for the 27B Clef ^[5]. At the median, that is about 5.4 times faster, or 170.5 ms saved per call ^[13]. Michelle Chen, Alex Reneau and Kevin Flansburg of Cloudflare wrote that because the models run on the company's edge GPUs, "you could put Clef into the hot path for agents to make decisions and combine that with one of our LLMs on Workers AI to take action" ^[15].

The 38.8 ms figure transfers only if three things hold. The call has to run on Cloudflare's hosting. Your state and schema have to be about the size of the benchmark's. And the median has to be the statistic your agent cares about. InfoQ's account reports medians only and does not describe the benchmark inputs ^[5]. Teams that download the weights from Hugging Face ^[9] run on their own hardware, and the edge-latency argument stays with Cloudflare ^[15].

The harder question is whether the probabilities mean anything. A per-option score lets an agent set a threshold below which it defers to a human ^[3]. That works only if the score drops when the model is wrong. On Reddit, bugra_sa wrote: "I'd care more about whether Clef knows when to punt than its raw accuracy score." The same user proposed testing on cases "where a false positive is much more expensive than a miss," then changing the data enough "to see when its confidence falls apart" ^[7]. On Hacker News, SebastianSosa wrote: "Public benchmarks are easy to cheat" ^[6].

Customers adapt Clef to their own data through a fine-tuning service that starts with help from Cloudflare engineers. A self-service platform is planned, and no firm date has been announced ^[8]. Cloudflare's own first targets are support triage and bot classification, tuned on its historical labelled data ^[12].