Cloudflare Clef is an open-weight swing at Jev, on a vendor scoreboard

Cloudflare released Cloudflare Clef and Clef-flash on October 1, 2026, on Workers AI, and put the weights on Hugging Face under Apache 2.0. The Register's story is timestamped 21:39 UTC. The blog and the Register are describing one launch, then parting ways on how to read Jev's context limit, what the hosted tokens cost, and how heavy a local run is.

TL;DR - On Cloudflare Clef, the blog says Clef's backbone is a frozen Qwen3.8-27B, Clef-flash is Qwen3.5-9B, both bring a vision encoder, and they are Jev-API compatible with a 64k context. - Self-reported median latency is 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev. The blog says Clef beat Jev on 3 of 4 Typesafe workflow tests. - The Register prices the hosted model at $0.24 per million tokens versus Jev at $0.042 per million, and it disagrees with the blog's flat "Jev's 32k" line.

What actually happened

The benchmark table is vendor-scored. Read it that way.

On Cloudflare Clef, this is a decision-model launch sitting on top of existing Qwen backbones, not a claim that Cloudflare trained a giant base model from a blank page. The blog says Clef's backbone is frozen Qwen3.8-27B and Clef-flash's is Qwen3.5-9B. There is a vision encoder in the picture.

Context is given as 64k, set against what the blog calls Jev's 32k. The API story is compatibility with Jev's API, which is the practical hook if you already speak that interface.

The latency numbers are self-reported medians. Clef comes in at 209.3 ms, Clef-flash at 38.8 ms, and Jev at 524.1 ms in that comparison. A separate domain-classification example puts Browser Run at 2.2 seconds against 4.7 seconds for gpt-oss-120b.

The blog also says Clef beat Jev on 3 of 4 Typesafe workflow tests. None of that is an outside lab rerunning the table. Vendor-scored means the company throwing the party also held the stopwatch.

On Cloudflare Clef, the Register adds the parts a speed table likes to skip. Hosted price is $0.24 per million tokens, against Jev at $0.042 per million. Faster in the vendor's median latency column and more expensive in the Register's hosted price column can both be printed on the same day.

They answer different questions. A procurement note that only copies the milliseconds is incomplete, and a note that only copies the token price is incomplete too.

Michelle Chen, in the Register's account, said the training data is not public. That is a sourcing fact, not a colorful aside. You can fetch weights under Apache 2.0 and still not know what the model was trained on.

Open weights and open data are different sentences. This launch gives you the first and, by Chen's statement, not the second.

Local hardware is the other cold shower. The Register says Clef-flash needs 41 GB of VRAM and Clef needs 85 GB, at single concurrency and a 64k context. That is not a laptop footnote. If "open weights" made someone on your team picture a overnight download and a consumer GPU, these figures say otherwise for a full-context, single-concurrency setup.

What to do as a reader

Split the scoreboard before you brief anyone. Latency medians, the Typesafe 3-of-4 line, the 64k context, the frozen Qwen backbones, the vision encoder, Jev-API compatibility, and the domain-classification example come from Cloudflare's blog. Hosted price, the VRAM figures, and Chen's line on training data come from The Register.

The comparison table for speed is vendor-scored even when a newsroom repeats numbers from it. Repetition is not a second stopwatch.

On Cloudflare Clef, hold the context-window disagreement in plain sight. The blog sets a 64k context against "Jev's 32k." The Register says Jev can take 64k tokens across a request, while state plus the longest question is capped at 32k. That is narrower than the blog's phrase, and the two write-ups disagree.

Do not "resolve" it by picking the version that makes a chart look cleaner. If your integration depends on how much state rides along with the longest question, the Register's wording is the one that changes the design, and it is still one outlet's reading.

Use the price and the speed as a pair. A self-reported median of 38.8 ms or 209.3 ms does not cancel a hosted rate of $0.24 per million tokens, and that rate does not cancel the latency claim. Jev's $0.042 per million in the Register is a different product's hosted price, useful only as the comparison they printed.

It is not a suggestion about which bill you should want. This is a reader note, not a purchasing instruction.

On Cloudflare Clef, if you are thinking about running weights yourself, start from the VRAM line and the license line, not from the latency line. Apache 2.0 on Hugging Face is the sharing term they announced. 41 GB and 85 GB are the Register's figures for Clef-flash and Clef at single concurrency and 64k context.

Training data staying private is Chen's statement via the Register. A demo that feels quick on Workers AI does not tell you that you can reproduce the medians on hardware you have not measured.

What the pages still do not do

They do not hand you an independent rerun of the four Typesafe tests. They do not publish the training set. They do not, in this pack, give a full price card beyond the Register's per-million hosted figure, and they do not itemize what Browser Run includes beyond the name in that timing example. The 2.2 second versus 4.7 second comparison is one example, against gpt-oss-120b, not a general proof about every routing task.

Jev-API compatibility is a migration claim from the blog. Compatibility is not identity. A frozen backbone plus a vision encoder can speak a familiar API and still behave differently on the workflows you care about.

On Cloudflare Clef, the honest pilot is a handful of your own decision traces, timed on your side, with the context shape written down. If your requests look like the Register's "state plus longest question" case, write that down too, because that is where the 32k disagreement actually bites.

The two pages to open are Cloudflare's Clef post and The Register's write-up. Keep the vendor table in the vendor's column. Where the context cap is concerned, say they disagree and move on.

Readers who want sourced recaps that are already on the site can read DogeOS public testnet and PUMP token buyback burn as separate live posts.