Cloudflare released two open-source decision models, Clef and Clef Flash, on October 1, 2026, alongside a new reinforcement learning (RL) fine-tuning platform. These models are designed to provide typed, probability-scored answers against a defined schema, differentiating them from traditional language models that generate free-form text. The announcement, published on Cloudflare's blog, details the models' performance, architecture, and potential applications.

Performance Metrics

The released models demonstrate significant latency improvements compared to existing solutions. Clef achieves a median latency of 209.3 milliseconds, while Clef Flash records a median latency of 38.8 milliseconds. For comparison, Jev from Typesafe AI has a median latency of 524.1 milliseconds. At the P95 metric, Clef's latency is 238.6 milliseconds, and Clef Flash's is 122.4 milliseconds. Cloudflare highlights that Clef Flash's P95 latency is roughly triple its median, but still under an eighth of a second, indicating a predictable performance band. This speed advantage allows for decision models to be integrated into loops within agent workflows, whereas slower models necessitate user-facing spinners.

Architectural Innovations

Clef and Clef Flash are not trained from scratch. Cloudflare states they utilize QN as the base model, post-trained for decision use cases. QN3A-27B is frozen for Clef, and QN35-9B is frozen for Clef Flash. This freezing process means the backbone weights remain unchanged. The training jointly optimized a routing head alongside rank 256 low-rank adapters.

A key architectural feature contributing to Clef's speed is its inference process. It performs a prefill-only pass with QN, followed by scoring valid schema choices in parallel. Unlike autoregressive LLMs that generate text token by token, Clef skips the decode phase entirely. Cloudflare describes this as a non-autoregressive step, significantly accelerating the decision-making process. The architecture is characterized as a two-stage attention routing process, combining option-specific evidence routing, joint cross-field attention, and schema-bound scoring with a lexical prior. Currently, no detailed paper or reproducible details have been published alongside this description.

Calibration and Optimization

The probability scores provided by Clef are calibrated through post-training using label-smoothed cross-entropy and Brier loss. Cross-entropy trains the model to select the correct answer, while Brier loss ensures the model is honest about its confidence levels. Cloudflare also developed Reinforcement Learning for Calibrated Decisions (RLCD) as a secondary optimization target. RLCD provides partial credit to adjacent ordinal choices, rewards precise record outputs, and applies a penalty to prevent distribution shifts. This approach ensures that confidence curves are meaningful, treating near misses as such rather than complete failures.

Evaluation Benchmarks

Cloudflare conducted 43 evaluation benchmarks. On the Berkeley Function Calling Leaderboard (BFCL) exact match, Clef achieved 98.47% and Clef Flash 98.76%, compared to Jev's 95.75%. On tool NDCG@10, Clef scored 69.19% and Clef Flash 66.43%, against Jev's 65.28%. For API bank accuracy, Clef reached 91.93% and Clef Flash 93.11%, with Jev at 88.19%.

In the home appliances case exact match, Clef achieved 82.95% and Clef Flash 97.73%, significantly outperforming Jev (52.27%), Diffusion Gemma (42.05%), and Kev 9B (25.00%). On banking 77 macro F1, Clef scored 94.20% and Clef Flash 90.93%, compared to Jev's 79.74%. On CLNC50+OS macro F1, Clef achieved 97.43%, while Clef Flash scored 66.77% and Jev 89.27%. Cloudflare notes that on CLNC50+OS, the smaller model's performance drops significantly, while the larger model achieves the top score.

Cloudflare also acknowledges benchmarks where Jev leads: when-to-call accuracy (Jev 80.97% vs. Clef 72.37%) and Bright NDCG@10 (Jev 47.52% vs. Clef 45.91%). Diffusion Gemma leads fish and chips accuracy at 85.35% against Clef's 79.60%.

On Typesafe's own evaluation suite, Cloudflare states Clef won three out of four areas: invoice processing (64.7% vs. Jev 61.8%), customer service (76.3% vs. Jev 76.0%), and security incidents (62.9% vs. Jev 61.7%). Jev led agent trace observability at 71.6% against Clef's 68.5%. Cloudflare also claims Clef is currently the leader on the Jev decision index, with results available on a live demo site.

Feature Parity and API Compatibility

Clef offers a 64K context window, double that of Jev's 32K. Additionally, Clef includes a vision encoder capable of classifying visual content, a feature stated to be absent in Jev's current text-only classification capabilities. The models are fully Jev API compatible, allowing for integration as a configuration change rather than a complete rewrite. The API call involves an HTTP POST request to api.cloudflare.com/client/v4/accounts/{your_cloudflare_account_id}/run/cf/cloudflare/clef. The request body includes a state field and questions, which can be of types null, choice with named criteria, or score with an ordered criteria list.

Internal Testing and Use Cases

Cloudflare's threat intelligence team utilized Clef Flash for website domain classification within a workflow called "browser run." In this test, Clef Flash took 2.2 seconds to fetch, render, and classify a website. For comparison, GPT-O120B, described as Cloudflare's fastest general LLM, took 4.7 seconds and returned only two classifications. This internal test, while not a published benchmark, suggests a potential for significant latency and result savings. Cloudflare frames this as a roughly 2x saving in both latency and the number of classifications returned.

Fine-Tuning Platform

Cloudflare has introduced a reinforcement learning fine-tuning platform designed to enhance model performance for specific tasks. The pipeline includes five components: AI Gateway for data capture from traffic, Workers AI for generating rollouts against base Clef, Cloudflare containers as an RL sandbox, a new trainer component for weight updates, and Workers AI plus BYO (Bring Your Own) model redeploys. The BYO model work is labeled COG, stemming from Cloudflare's acquisition of Replicate.

Potential internal use cases for this fine-tuning include evaluating trust and safety submissions, triaging support requests, and classifying bots in their bot products. Cloudflare cites over 15 years of network data as fine-tuning material. The company acknowledges that fine-tuning may reduce general-purpose performance in favor of higher accuracy within a specific domain. The claim that a fine-tuned Clef is more accurate and faster than the generic model is asserted without supporting numbers.

Open Source and Availability

Both Clef and Clef Flash models are open-sourced on HuggingFace under the Apache 2.0 license, permitting modification and redistribution for commercial use. The weights are available for local execution. The hosted endpoint is documented on the Workers AI platform, which provides GPU inference across Cloudflare's global edge. Cloudflare has also implemented a policy stating that customer requests or responses are not read, stored, or trained on unless the fine-tuning product is used.

Unanswered Questions

Several details remain unpublished, including pricing for the models and the self-served platform, the general availability status of both models, specific parameter counts and hardware requirements for local deployment, the exact ship date for the trainer component, whether vision capabilities are live in the hosted API, image limits, rate limits, and the maximum number of questions per call. The specifics of the "browser run" workflow and customer ownership of weights trained by Cloudflare's FTE team are also not detailed. Cloudflare is inviting existing customers to participate as design partners for the fine-tuning and RL work, suggesting that some of these answers will emerge from that cohort.

Target Audience and Market Impact

Clef is positioned for use cases involving agents that currently rely on expensive general-purpose reasoning models for tasks such as routing, triage, moderation, domain classification, and severity scoring. The trend towards smaller, specialized classifiers replacing costly general calls is noted. Cloudflare's release of a calibrated, typed, API-compatible model with open-source weights is seen as a significant development in this area. Clef is described as Cloudflare's first Cloudflare-trained ML model from the Workers AI team, with the potential to disrupt agent utilization and align with Cloudflare's mission as an "agent cloud." The name "Clef" is derived from music theory, signifying the assignment of pitch names to a staff, analogous to the model's function of turning raw context into labeled, scored meaning.

The article concludes by emphasizing that Clef represents a decision model where speed, calibration methods, and licensing align. The median latency of 38.8 milliseconds, Brier calibration, and Apache 2.0 license with open weights are highlighted. Readers are encouraged to verify the published numbers through independent reruns on HuggingFace and the Workers AI platform.