Cloudflare’s Clef and Clef-flash are substantial Jev alternatives: 27B and 9B multimodal decision models with Apache-2.0 weights, a Jev/SystemOne-compatible API, inference code, and a published 43-benchmark comparison. Unlike Jev, both models can classify text, JSON, images, and video.

I inspected Cloudflare’s announcement and both Hugging Face model cards. I did not download the 9B or 27B weights, run inference, or reproduce the benchmark. Every quality and latency number below is Cloudflare’s measurement.

What Cloudflare actually released

  • Clef starts from Qwen3.8-27B; Clef-flash starts from Qwen3.5-9B.
  • Each release includes sharded backbone weights, a joint schema head, tokenizer and media processor files, and joint_schema_model.py for loading, batching, and Jev-compatible responses.
  • The model performs a prefill-only backbone pass, then scores every valid answer in parallel instead of generating an answer token by token.
  • A small transformer head routes evidence from the state to each question and scores the questions jointly.
  • Both models accept Noul, Choice, and Score questions and return one probability per allowed option.
  • The hosted Workers AI endpoint is intended as a drop-in Jev/SystemOne target; the downloadable models are Apache-2.0 licensed.

This is more open than a hosted API wrapper: the released package includes the model weights and the code that maps a typed request to probabilities. Cloudflare does not publish the full training dataset, however, so “open” does not make the complete training run reproducible.

The benchmark is broad—and self-run

Across Cloudflare’s internal run of Decision Index 0.2.1:

  • Clef led Jev on BANKING77 macro-F1, 94.2% versus 79.7%.
  • Clef-flash led on API-Bank accuracy, 93.1% versus 88.2%.
  • Jev remained well ahead on GPQA Diamond, 78.3% versus 48.0% for Clef and 51.0% for Clef-flash.
  • Jev also led on BRIGHT, MMLU-Pro, BBH, HoVer, When2Call, and several other tasks.
  • DiffusionGemma, not Clef or Jev, led the published PhishNChips result at 85.4%.

On TypeSafe’s four workflow families, Cloudflare reports Clef or Clef-flash ahead in invoice processing, customer service, and security incidents, while Jev led agent-trace observability. The exact-action scores were often only in the low 60s or 70s. Typed output makes those responses structurally valid; it does not make the selected actions correct.

Cloudflare reports median request latency of 209.3 ms for Clef, 38.8 ms for Clef-flash, and 524.1 ms for Jev across the suite. Laya had the lowest median at 5.8 ms. These are Cloudflare’s hosted measurements, not a matched independent hardware test, and network placement is part of the result.

The architecture is the interesting contribution

  • The frozen Qwen backbone is adapted with rank-256 LoRA modules.
  • A joint schema head reads final hidden states and produces one logit for every allowed option.
  • Cloudflare says training combines label-smoothed cross-entropy with Brier loss for probability calibration.
  • Its RLCD stage gives partial credit to nearby ordinal answers, rewards exact record-level outputs, and uses a reference penalty to limit distribution shift.
  • Clef retains Qwen’s vision encoder, extending the decision interface beyond Jev’s text-only input.

That architecture occupies useful middle ground: more flexible than a classifier trained for one fixed label set, but narrower and potentially cheaper than asking a generative model to reason and emit JSON at every branch.

The operational caveats stay the same

  • A legal option with a concentrated probability distribution can still be the wrong decision.
  • Cloudflare’s benchmark used Cloudflare’s harness and infrastructure; an independent rerun is needed before treating the ranking as settled.
  • The 27B model card says local use was tested on a single H200, so “downloadable” does not mean lightweight.
  • Clef’s downloadable encoder defaults to 16,384 tokens even though the hosted product advertises a 64k context window; deployments should verify which limit applies.
  • Vision widens the attack and evaluation surface. Image or video support does not establish robustness to adversarial media.
  • Fine-tuning can improve one domain while reducing general performance. Thresholds and calibration need a held-out set from the actual workflow.

Takeaway

Clef matters because Cloudflare released an inspectable model family, not merely another hosted Jev clone. The weights, joint schema head, inference path, multimodal inputs, and compatible API make it possible to study or replace the decision layer without changing the surrounding application contract.

The benchmark supports a narrower conclusion than “Clef beats Jev”: Cloudflare’s models are competitive across a wide suite, often faster in Cloudflare’s measurements, and uneven by task. The right production comparison is still a fixed local dataset, a fixed decision policy, and measured accepted-case error—not the model with the most bold cells in its own table.