Cloudflare’s Clef puts the decision machine on your workbench

The lovely thing about a new AI category is how quickly somebody arrives with a screwdriver.

TypeSafe introduced Jev in early access on September 15. Cloudflare announced Clef and Clef-flash on October 1, sixteen days later. Both Cloudflare models have downloadable weights under Apache 2.0 and a Jev-compatible API. Viddy the workbench, droogies. There are more parts to choose from already.^[1][2][3][4]

Take the casing off

Clef is the 27-billion-parameter model; Clef-flash is the 9-billion-parameter version. Their model cards describe a machine that takes some state and a set of typed questions, then scores the permitted answers in one forward pass. It can work with text and images. The output is probabilities over choices, rather than a little essay your software has to interpret.^[3][4]

That could mean sorting a support ticket, choosing a team or assessing urgency. Useful, gloriously unromantic work. Sometimes the computer just needs to pick a drawer without writing a memoir about the drawer.

The practical change is that developers can inspect, adapt and run these weights themselves, as well as use Cloudflare’s hosted service. API compatibility gives people already using Jev’s request shape another implementation to evaluate. Compatibility alone says nothing about identical answers, training history or who deserves the conceptual crown.^[1][3][4]

Mind the demonstration goggles

Cloudflare reports median latency of 38.8 milliseconds for Clef-flash and 209.3 milliseconds for Clef, versus 524.1 milliseconds for Jev across its 43 evaluation benchmarks. Those are Cloudflare’s own results; your workload and network path still need testing. Its launch table also shows Jev ahead of both Clef variants on When2Call and BRIGHT. The victory parade should probably keep the receipt.^[1]

A permitted answer can still be the wrong answer. If software is going to act on these scores, our preference is boring: test on the actual workload, measure the expensive mistakes and leave an escalation path. A beautifully typed blunder remains a blunder, my brothers.

The weights travel. The meter still runs.

Cloudflare’s Workers AI pages list $0.09 per million input tokens for Clef-flash and $0.24 for Clef. Open weights do not make hosted inference free, and running them yourself still requires compute.^[5][6]

There’s also a services pitch. Cloudflare is offering hands-on fine-tuning with its forward-deployed engineering team; a self-serve platform is planned for later. Read that tense carefully before adding a nonexistent button to your project plan.^[1]

Our read: the interesting contest is moving into the dull, consequential plumbing of software. Cheap decisions, swappable interfaces, weights you can take home. Less theatrical chattering. More scrutiny of what happens when the machine says yes.