Cloudflare Clef: Open Decision Model Tested on RTX 4090

👁 114 reads

On October 1, 2026, Cloudflare released two open decision models: Clef (27B) and Clef-flash (9B). They do not generate running text - they take a context state and a list of questions with permissible options, returning the calibrated probability distribution for each option. Weights are available on Hugging Face under the Apache-2.0 license, and the hosted service on Workers AI starts at $0.038 per million tokens (Clef-Flash, following a price cut on October 9). I ran Clef-flash locally on an RTX 4090 - on 30 customer tickets tested in Polish, it routed to the correct department in 30 out of 30 cases with a median latency of 112 ms. I cover the broader model category and comparisons in my guide to AI decision models.

What Cloudflare Actually Released

• Clef - 27B parameters, built on top of a frozen Qwen3.8-27B. Weights take up 55 GB in bf16.
• Clef-flash - 9B parameters, built on Qwen3.5-9B, with 19 GB of weights. Designed for low-latency decision routing.
• Apache-2.0 license for both, matching the base Qwen models. Free for commercial deployment.
• Hosting on Workers AI under model IDs @cf/cloudflare/clef and @cf/cloudflare/clef-flash.
• Fine-tuning service (RL fine-tuning) - initially as an engagement with Cloudflare’s engineering team, later shifting to a self-serve platform.

The Cloudflare announcement was authored by Michelle Chen, Alex Reneau, and Kevin Flansburg. The name refers to the musical clef that defines note pitches on a stave - with “CF” doubles as Cloudflare’s initials.

How the Joint Schema Head Works

A standard language model, even when constrained to JSON outputs, decodes token by token. Clef avoids that overhead completely:

1. It processes the input context (text, JSON, images, or video) through the base Qwen model only once during the prefill phase.
2. Hidden representations are fed into a compact head that Cloudflare calls the joint schema head. In Clef-flash, this module features 4 layers with a width of 1024.
3. For each candidate option, the head extracts matching elements from the input, allows questions to attend across one another, and only then scores the candidates.
4. The final output is a single logit per choice. Softmax within each question produces the probability distribution.

Cloudflare trained the head along with rank-256 LoRA adapters while keeping the core Qwen base frozen. Training incorporated Brier scores into the loss function to penalize uncalibrated confidence, paired with an RLCD (Reinforcement Learning for Calibrated Decisions) second phase that awards partial credit for near-miss scale selections. Training datasets were Cloudflare’s internal synthetic corpora.

Request Format: Compatible with Jev

Clef uses the exact same API format introduced by TypeSafe’s Jev on September 15, which established the decision-model category (covered in my analysis of TypeSafe and the Jev model). The payload requires a state field and a questions dictionary supporting three types: choice (categorical pick), noul (yes/no boolean), and score (numerical scale). Cloudflare designed it as a drop-in replacement, allowing software written for System One to migrate by changing the endpoint URL. Jared Palmer’s Kev and AWS’s Strands Decider - launched on the same day as Clef - follow this exact specification.

Cloudflare brings two proprietary extensions: an image encoder (enabling Clef to verify if an uploaded receipt is readable, for instance) and a 64k token context window. Workers AI imposes specific limits: 1 to 64 questions per call, up to 4 images at 4 MiB each, and an overall request ceiling of 13 MiB.

Official Cloudflare Benchmark Numbers

Cloudflare published comprehensive tables covering 43 evaluation suites alongside comparisons against competing decision models on their model card and Decision Index pages. CLINC150 deserves special attention: the smaller Clef-flash exhibits a severe drop compared to the 27B model, showing that “flash” is not simply an identical drop-in substitute at lower cost.

In production, Cloudflare’s Threat Intelligence group uses the model for domain classification. Clef completed page retrieval, rendering, and classification in 2.2 seconds, whereas gpt-oss-120b took 4.7 seconds on the identical workload - and only returned two top categories instead of a complete probability spread.

Pricing Breakdown

You pay solely for input tokens because the model produces no generated text output. The daily grant of 10,000 Neurons applies across both Workers Free and Paid plans; usage past the quota costs $0.011 per 1,000 Neurons. Workers AI requires an active Cloudflare account - for this evaluation I ran the model locally, so hosted observations rely on documentation. You can read more about the platform’s infrastructure in my guide to Cloudflare.

Local Testing: Clef-flash on an RTX 4090

• Download: 19 GB from the official Cloudflare/clef-flash repository. Loading from a slow external USB drive took approximately 7.5 minutes.
• Memory consumption: Peaked at 18.3 GB VRAM, fitting comfortably inside the 24 GB card limit.
• Evaluation results: Tested across 30 customer support tickets in Polish covering billing, shipping, returns, account access, outages, and sales inquiries. Department categorization scored 30 out of 30, with a median turnaround of 112 ms for a two-question evaluation.
• Local comparison: Gemma 4 12B in Ollama using constrained JSON outputs also scored 30/30, but took 413 ms, while Qwen3.8 27B clocked in at 762 ms.
• Threshold sensitivity: On an “is this urgent?” classification, Clef-flash generated false positives at a default 0.5 decision threshold. Raising the cutoff to 0.95 dropped the error rate to 2 mistakes out of 30. You will need to calibrate this threshold on your own dataset.

I did not run the full 27B Clef locally: 55 GB in bf16 cannot fit into a 24 GB GPU, there is no official quantized release, and the custom joint schema head rules out standard Ollama and GGUF runtimes. Full benchmarks spanning eight models, performance graphs, and architectural conclusions are compiled in my guide to AI decision models. To inspect techniques for extracting constrained outputs from standard LLMs, refer to my walkthrough on structured output.

Why Decision Models Matter

Within three weeks, decision models moved from a proprietary API novelty to multiple open weights running on shared schemas: Kev, Clef, Strands Decider, and OpenAI’s own Decisions API previewed at DevDay (covered in my writeup on the Agents API). If you build AI agents or automate support ticket triage, offloading small branching decisions - routing to departments, escalation flags, tool dispatching - away from massive generative LLMs makes workflows dramatically faster, cheaper, and statistically calibrated. Cloudflare complements this with edge deployment and custom fine-tuning, pitching a unified stack for enterprise routing pipelines.

Sources

• Artificial Analysis Decision Index v1.0 (October 13, 2026)
• Workers AI Changelog (Clef-Omni, updated Clef-Flash pricing, Oct 9, 2026)
• Cloudflare Blog: Introducing Clef (October 1, 2026)
• Clef-flash Model Card and Clef Model Card
• Workers AI Pricing
• Local test environment: October 5, 2026, RTX 4090, PyTorch 2.11, transformers 5.18.

// read moreAI Decision Models: Decision Index BenchmarkBenchmarksAWS Strands Decider 2B Tested: Fast Open Source Decision ModelBenchmarksQwen-Image-2.1 vs Closed Models: 4 Tests on RTX 4090Benchmarks

Kinetyra