# chen-reneau-flansburg-clef-decision-models-2026-10-01

## Veille

Post by **Michelle Chen, Alex Reneau and Kevin Flansburg**, published on **October 1, 2026** on the **Cloudflare** blog (~10-minute read, "AI Birthday" week). Cloudflare releases **Clef** and **Clef-flash**, its first in-house trained models, hosted on **Workers AI** and distributed as open source (Apache 2.0, Hugging Face), along with a reinforcement-learning fine-tuning service.

## Titre Article

Introducing Clef: our open-source decision models, and new RL fine-tuning platform

## Date

2026-10-01

## URL

https://blog.cloudflare.com/clef-decision-models/

## Keywords

Clef, Clef-flash, decision model, Cloudflare, Workers AI, Jev, Jev Decision Index, open source, Apache 2.0, fine-tuning, reinforcement learning, RLCD, Qwen, DiffusionGemma, logprobs, calibration, latency, classification, AI Gateway, cloud agent

## Authors

Michelle Chen, Alex Reneau, Kevin Flansburg — équipe Workers AI / plateforme IA, blog Cloudflare.

## Ton

Product-launch post from an engineering team, positive and technical in tone. The opening positions Clef against the "buzz" around Jev, then alternates concrete examples (domain classification by Threat Intelligence, support ticket triage), benchmark tables, and training details (rank-256 adapters, Brier loss). The name is explained through a musical analogy (the clef that fixes the staff's note range; "CF" for Cloudflare). The text speaks of the platform's future: the fine-tuning service starts with forward-deployed engineers (FDE), the self-service version comes later, and Cloudflare is looking for design-partner customers.

## Pense-betes

- **Definition stated in the article**: a decision model returns typed answers with probabilities; the code routes, escalates, or hands off to a human. Framework: agents that gather context, decide, and act, with a human only when necessary.
- **Internal case**: domain classification by the Threat Intelligence team with Browser Run: **2.2 s** to fetch, render, and classify, versus 4.7 s for gpt-oss-120b, which returned only two categories.
- **Selected results** (Clef / Clef-flash / Jev): BFCL 98.47 / 98.76 / 95.75; BANKING77 94.20 / 90.93 / 79.74; When2Call 72.37 / 65.58 / 80.97; BRIGHT 45.91 / 39.26 / 47.52. On TypeSafe's workflows: billing 64.7 vs 61.8, agent trace observability 68.5 vs 71.6.
- **Training**: frozen Qwen base (Qwen3.8-27B for Clef, Qwen3.5-9B for Clef-flash), routing head and rank-256 adapters, label-smoothing cross-entropy plus Brier loss, then **RLCD** with partial credit for neighboring ordinal choices. The initial approach adapted DiffusionGemma by exposing logprobs.
- **RL service**: AI Gateway (dataset from traffic), Workers AI (rollouts), Containers (scoring sandbox), *Trainer* (new), Workers AI with a brought model (Cog, inherited from Replicate). Customer data is not read or stored, except for fine-tuning.
- ⚠️ **Caveats visible in the tables**: Clef-flash drops to 66.77 on CLINC150+OOS and Jev stays ahead on When2Call, BRIGHT, and PhishNChips (DiffusionGemma Jev at 85.35); Laya is faster but notably less accurate.

## RésuméDe400mots

On **October 1, 2026**, **Michelle Chen, Alex Reneau and Kevin Flansburg** announce on the **Cloudflare** blog two decision models, **Clef** and **Clef-flash**, hosted on **Workers AI** and published on Hugging Face under the Apache 2.0 license, along with a reinforcement-learning fine-tuning product. The post opens from the "buzz" around TypeSafe's **Jev** model: a decision model produced cheaply, quickly, and consistently returns bounded structured outputs, meant to be inserted into a workflow when a decision is required, as opposed to LLMs, which are largely non-deterministic but open-ended.

A decision model receives a state, for example a support message, and answers typed questions (urgency, team, severity) with probabilities, which the code uses to route, escalate, or hand off to a human. At Cloudflare, Clef classifies domains for Threat Intelligence: 2.2 seconds to fetch, render, and classify a site, versus 4.7 seconds for gpt-oss-120b, which returned only two categories.

Clef differs from Jev through a **vision encoder**, hence image classification, and a **64k** context versus 32k. The API is compatible with Jev's, making replacement straightforward. On the *Jev Decision Index*, Clef leads; the tables show clear advantages (BFCL 98.47, BANKING77 94.20, CLINC150+OOS 97.43) and cases where Jev or a variant stay ahead (When2Call, BRIGHT, PhishNChips). On TypeSafe's workflow evaluations, Clef beats Jev on three of four. Median latency: 209.3 ms for Clef, 38.8 ms for Clef-flash, 524.1 ms for Jev; hosting on the edge GPU network reduces network latency. Clef can thus sit on an agent's critical path, ahead of an LLM tasked with acting.

Training: the project started from an adaptation of DiffusionGemma exposing logprobs, building on work by Matt Mastracci. Clef uses a frozen Qwen3.8-27B (Clef-flash: Qwen3.5-9B) for a prefill-only pass, then scores the schema's valid choices in parallel, without text generation. A two-stage attention routing head and rank-256 adapters are trained with label-smoothing cross-entropy and a Brier loss for calibration, then with **RLCD**, which grants partial credit to adjacent ordinal choices.

The fine-tuning service initially targets customers supported by forward-deployed engineers, with a self-service version to follow. It relies on AI Gateway (dataset), Workers AI (rollouts), Containers (scoring sandbox), a new *Trainer* (weight updates), and deployment of brought models. Cloudflare sees this as a way to specialize Clef (moderation, support triage, bots) using its fifteen years of labeled data, and to embody its "cloud agent."

## GrapheDeConnaissance

- Michelle Chen —travaille_chez→ Cloudflare (ORGANISATION, 0.9)
- Cloudflare —publie→ Clef (TECHNOLOGIE, 0.98)
- Cloudflare —publie→ Clef-flash (TECHNOLOGIE, 0.98)
- Clef-flash —est_variante_de→ Clef (TECHNOLOGIE, 0.93)
- Clef —est_instance_de→ modèle de décision (CONCEPT, 0.97)
- Clef —converge_avec→ Jev (TECHNOLOGIE, 0.92)
- Clef —surpasse→ Jev (TECHNOLOGIE, 0.85)
- Clef —est_basé_sur→ Qwen (TECHNOLOGIE, 0.96)
- Clef —utilise→ RLCD (METHODOLOGIE, 0.95)
- Clef —s_inspire_de→ DiffusionGemma (TECHNOLOGIE, 0.88)
- Clef —mesure→ latence médiane de 209,3 ms (Clef-flash : 38,8 ms) contre 524,1 ms pour Jev (MESURE, 0.95)
- Clef —mesure→ 2,2 s pour classer un domaine contre 4,7 s pour gpt-oss-120b (MESURE, 0.95)
- Clef —utilise→ Workers AI (TECHNOLOGIE, 0.96)
- service de fine-tuning RL —utilise→ AI Gateway (TECHNOLOGIE, 0.94)
- service de fine-tuning RL —utilise→ Cloudflare Containers (TECHNOLOGIE, 0.94)
- Cloudflare —publie→ service de fine-tuning RL (TECHNOLOGIE, 0.95)
- Cloudflare —affirme_que→ Clef a le potentiel de changer la façon d'utiliser les agents (AFFIRMATION, 0.8)
- Jev —permet→ décisions typées à probabilités calibrées (CONCEPT, 0.85)

---
Canonical: https://www.thekb.eu/en/fiches/chen-reneau-flansburg-clef-decision-models-2026-10-01/