Cloudflare ships Clef, open source decision models

Cloudflare released Clef and Clef-flash, open source decision models hosted on Workers AI, along with a new reinforcement learning fine-tuning platform. Cloudflare says Clef currently leads the Jev Decision Index, and the models are fully API-compatible with TypeSafe’s Jev, so swapping them in does not require rewriting the call sites.

A decision model here means what it says: a classifier that returns bounded typed outputs with probabilities rather than free text, so code can route, escalate or defer to a human. The example Cloudflare gives is a support message classified for urgency, team and severity. The models are Apache 2.0 on Hugging Face for local runs, and the API identifier is @cf/cloudflare/clef.

Two things set Clef apart from Jev as Cloudflare describes them. It has a vision encoder, so it can classify visual content, where Jev is text only today. And it has a 64K context window against Jev’s 32K, which lets you push more input state into a single classification.

Latency is the real headline

Across 43 evaluation benchmarks, median latency is 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev. p95 is 238.6 ms and 122.4 ms respectively. Cloudflare’s own example is concrete: classifying website domains for its Threat Intelligence team took 2.2 seconds with Clef plus Browser Run, against 4.7 seconds for gpt-oss-120b, which also returned only two classifications where Clef returned a full probability set.

On quality, Clef wins most of the Jev Decision Index benchmarks. BANKING77 macro-F1 is 94.20 against Jev’s 79.74, CLINC150 out-of-scope is 97.43 against 89.27, BFCL case exact is 98.47 against 95.75, and BFCL case exact for Clef-flash is 98.76. It is not a clean sweep: Jev leads When2Call at 80.97 against 72.37, BRIGHT at 47.52 against 45.91, and PhishNChips goes to DiffusionGemma Jev at 85.35 against 79.60. Across TypeSafe’s own WorkflowEvals suite, Clef beats Jev in three of four areas, losing agent trace observability 68.5 to 71.6.

The training approach is the other interesting part. Clef freezes a Qwen backbone, Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash, and optimizes a routing head alongside rank-256 LoRA adapters. At inference it runs a prefill-only pass and then scores valid schema choices in parallel, so the decision step is non-autoregressive with no intermediate text. Loss is label-smoothed cross-entropy paired with a Brier loss for probability calibration, plus what Cloudflare calls RLCD, which gives partial credit to adjacent ordinal choices and applies a reference penalty against distribution shift.

One caveat on the latency numbers: Laya is faster still at 5.8 ms median, Cloudflare notes, though it trades heavily on the quality benchmarks. The RL fine-tuning service is also not a self-serve product yet. Cloudflare is running it as a hands-on engagement with its forward-deployed engineer team and looking for design partners before the self-serve platform opens up.

(Original source: Cloudflare)