Cloudflare ships two decision models, Clef and Clef-flash
Peter Steinberger flagged the release with one line and no test results.
Cloudflare has released two models it trained itself. The announcement line, quoted by Peter Steinberger, reads "Today, we're releasing two Cloudflare-trained decision models, Clef and Clef-flash". Steinberger's entire reaction was one sentence about how quickly the idea is travelling.
That is the whole of what is on the record so far from the people we follow. Steinberger did not say he had run either model, did not post latency or accuracy numbers, and did not compare them to anything else. The naming pattern, a model plus a cheaper and faster sibling, is the only hint in his post about how the two are meant to be used, and he does not spell out which workloads they are for.
What is actually new here
The news is the identity of the vendor, not a score. An infrastructure company shipping its own trained decision models is a company deciding that the small model that makes a cheap call, rather than the large model that writes the code, is something it would rather own than buy. Cloudflare already sells the pipes those calls travel through.
Steinberger's line is a read on momentum rather than on quality.
For working engineers, treat this as a release note and not as a result. Nobody in the sources has benchmarked Clef or Clef-flash, so there is no basis yet for saying whether they beat whatever small model you are currently calling in front of your expensive one. The useful question when numbers do appear is the same one that applies to every model in this slot, which is cost and latency per decision at a quality you can live with, measured on your own traffic rather than on a published table.
If you are already running a cheap model to classify, triage or route requests, the thing worth watching is whether running that step next to your edge compute, from the same vendor that serves the request, buys enough in round trip time to matter. That is an empirical question and the answer is not in yet.
Never seen an idea spreading so fast.