Cloudflare has released two open-weight decision models, Clef and Clef-flash, which handle structured choices and can also process images and video, according to a report carried by news.lavx.hu citing The Register. The launch gives developers another option beyond hosted decision APIs, although running the models locally comes with hefty memory demands.
Watch Desk analysis
What happened
The report says Clef and the smaller, faster Clef-flash answer bounded questions such as yes-or-no choices, multiple-choice questions and rankings. Cloudflare offers them through Workers AI and Hugging Face, and says the models use Qwen backbones. Their API is described as compatible with TypeSafe’s Jev, allowing developers to try Clef as a replacement without rebuilding an integration.
Cloudflare claims Clef outperforms Jev in three of four areas in tests using TypeSafe’s benchmarks. Its own results on the Jev Decision Index have not yet been reproduced for the official ranking, so the performance claims remain company-reported. The Register’s account says Clef costs US$0.24 per million tokens on Workers AI, compared with US$0.042 for Jev. For local use, the reported memory requirements are at least 85 GB of GPU VRAM for Clef and 41 GB for Clef-flash, assuming single concurrency and a 64k context window.
Key findings
- Two models target structured decisions
Clef is the larger model; Clef-flash is the smaller, faster option. - Image and video inputs are supported
That extends the models beyond text-only classification, according to the report. - The API is Jev-compatible
Developers can test Clef as a potential drop-in replacement for existing Jev integrations. - Weights are available to download
The report says the models are on Hugging Face under Apache-2.0 terms, while their training datasets are not public. - Local deployment has a substantial hardware bar
The reported minimums are 85 GB of GPU VRAM for Clef and 41 GB for Clef-flash.
Why it matters
Decision models turn a question into a structured answer rather than a paragraph, making them useful for tasks such as classification and routing. Clef’s reported image and video support broadens the kinds of inputs developers could test, while downloadable weights offer more control over deployment than a hosted-only service.
The trade-offs are concrete: the reported hosted price is higher than Jev’s, and local use calls for unusually large amounts of GPU memory. The benchmark picture is also not settled. Cloudflare’s comparisons are an interesting competitive signal, not an independently reproduced result.
Our read
Clef makes the decision-model category more practical to explore, particularly for teams that need image or video inputs or want downloadable weights. But “drop-in” compatibility does not make it a free upgrade: check the price, hardware requirements and the task-specific results before swapping out a working system. The leaderboard can wait for the receipts.
What to watch
- Whether Cloudflare’s Decision Index results are reproduced in the official ranking.
- How Clef and Clef-flash perform on real tasks involving images and video.
- Whether developers find the reported cost and hardware requirements worthwhile in practice.
Discussion spark: Would you choose an open-weight decision model with broader input support at a higher reported price, or stick with a cheaper option until its advantages are independently demonstrated?
Sources and evidence
- Cloudflare tries to outplay Jev with open-weight Clef models – news.lavx.hu (2 October 2026, 04:28 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.