Cloudflare Open-Sources Clef, a Direct Shot at TypeSafe AI's Jev — and the Ecosystem Already Has Receipts
On October 1, Cloudflare open-sourced Clef and Clef-flash, a pair of "decision models" built to answer bounded, structured questions — yes/no calls, multiple-choice routing, rubric scoring — rather than generate free text. Both ship under Apache 2.0 with weights on Hugging Face and hosted inference on Workers AI as @cf/cloudflare/clef and @cf/cloudflare/clef-flash. The move puts Cloudflare squarely after TypeSafe AI's Jev, the model that effectively created this decision-model category earlier in 2026 and against which Clef's API is deliberately compatible for drop-in swaps.
The architecture is the real story for anyone evaluating it as infrastructure rather than a chatbot replacement: Clef runs a prefill-only pass and then scores typed schema choices in parallel instead of autoregressively generating tokens, which is what lets it return calibrated probabilities in milliseconds. Clef is built on a Qwen3.8-27B backbone (85GB VRAM) and Clef-flash on Qwen3.5-9B (41GB VRAM), both with a 64k context window — double Jev's 32k — and, unlike Jev, both can classify images and video, not just text. Cloudflare's own Decision Index numbers put Clef-flash at roughly 39ms median latency versus Jev's 524ms, though that benchmark run is Cloudflare's own and, as of this week, hasn't been independently reproduced on the public leaderboard.
The ecosystem moved faster than the benchmarks got verified
What's notable is less the announcement than the 24-hour aftermath. Independent quantizers — bartowski, prithivMLmods, EliovpAI, mlx-community, and others — had GGUF, FP8, and MXFP4 builds of both Clef and Clef-flash live on Hugging Face within a day, the same pattern of near-instant community re-packaging that now greets any notable open-weight drop. More telling is that developers started wiring it into real projects almost immediately: a public GitHub pull request already shows a small open-source routing tool, fuzzy-decision, migrating its backend to Cloudflare's clef-flash model days after release — a faster real-world adoption signal than the download or trending-chart metrics the ecosystem usually leans on.
The pricing gap and the unverified claims
Independent reporting has flagged two caveats Cloudflare's own announcement understandably doesn't lead with. First, cost: Workers AI prices Clef around $0.24 per million tokens against Jev's roughly $0.042, close to a six-times premium that Cloudflare is betting its multimodal support and self-hosting option will justify. Second, and more fundamental, the head-to-head quality claims — that Clef beats Jev on three of four tested benchmark categories — come entirely from Cloudflare's internal Decision Index run; outside coverage has pointedly noted those results have not yet been reproduced on the index's official, third-party leaderboard, which is the standard this whole model category set for itself.
There's a sharper complication sitting underneath the pricing debate. A Check Point security study published just a week before Clef's launch found Jev's typed, non-autoregressive design offered essentially no protection against prompt injection: 25 of 27 attacks succeeded by feeding the model fabricated evidence embedded in normal-looking documents, at a cost of about $0.50 per successful manipulation, and neither structured input formatting nor explicit anti-injection instructions meaningfully helped. Because Clef shares the same non-autoregressive, typed-schema approach — and is pitched as a direct, API-compatible substitute for Jev — that finding raises an open question Cloudflare's launch materials don't address: whether scoring schema choices instead of generating tokens actually closes the manipulation gap, or just changes its shape.