I'm Nitin Kumar, Lead Project Coordinator at Debugged Pro. I work with software and IT agencies in India and the USA that ship agent features clients can audit. Cloudflare Clef landed on October 1, 2026 as Cloudflare's first Workers AI trained decision models.

Cloudflare published the launch on its official blog, with model pages on Workers AI and open weights on Hugging Face. The pitch is simple: agents need fast, bounded decisions, not another long chat reply.

This guide sticks to verified facts from Cloudflare and independent coverage in The Register. It maps those facts to delivery choices for India and USA agency teams that already run support, trust and safety, or tool-gated agents.

Cloudflare Clef decision models launch visual from the Cloudflare blog — Feature image: Cloudflare Clef launch artwork from the Cloudflare Blog post Introducing Clef. Credit: Cloudflare.

What Cloudflare Clef is and why agencies care

A decision model answers typed questions over an input state. It returns probabilities your code can branch on. Cloudflare frames Clef as the music clef that sets the domain for the notes that follow.

Cloudflare Clef ships as two hosted models: Clef for precision and Clef-flash for latency-critical paths. Both accept text, JSON, images, and video state. Both return structured answers your agent can act on without parsing prose.

For agencies, that matters when a client asks for triage, escalation, or a pre-tool safety check. Those steps need a clear schema, a probability threshold, and a human fallback, not a free-form essay.

Cloudflare Clef vs chat models in real delivery work

General LLMs can classify tickets. They also invent categories, stall on long reasoning, and force brittle parsers. Decision models constrain the output space up front.

Typed questions: noul for yes or no, choice for one option, score for an ordered rubric.
Up to 64 questions in one request, keyed by ids your code already understands.
No free-form text to scrape. Probabilities and chosen options come back ready for routing.
Vision support: pass up to four images with the state, which text-only classifiers still miss.

Cloudflare's Threat Intelligence example is concrete. With Browser Run, Clef fetched, rendered, and classified a domain in 2.2 seconds. The same workflow took 4.7 seconds on gpt-oss-120b and returned fewer categories.

Clef and Clef-flash: size, speed, and pricing

Pick the tier from the job, not from the demo. Cloudflare's changelog and docs publish the numbers agencies need for estimates.

Model sizes and context

Clef (@cf/cloudflare/clef): 27B parameters, 64K context, vision enabled.
Clef-flash (@cf/cloudflare/clef-flash): 9B parameters, same 64K context and vision support.
Backbones: Cloudflare says it froze Qwen3.8-27B for Clef and Qwen3.5-9B for Clef-flash, then post-trained for decisions.

Latency numbers Cloudflare reported

Across 43 eval runs, Cloudflare reported median latency of 209.3 ms for Clef and 38.8 ms for Clef-flash. Jev sat at 524.1 ms in the same table. p95 was 238.6 ms for Clef and 122.4 ms for Clef-flash.

Those figures are Cloudflare-reported. Still treat them as useful planning inputs for hot-path designs, then measure on your own traffic.

Hosted pricing from Workers AI docs

Clef: $0.24 per million input tokens.
Clef-flash: $0.09 per million input tokens.
Local option: Apache 2.0 weights on Hugging Face if you want to self-host.

The Register noted that hosted Clef costs more per token than Jev at about $0.042 per million. Read their Clef coverage before you promise a straight cost win. Speed, vision, and edge hosting may still justify the premium for clients who care about latency.

The Register coverage image for Cloudflare Clef decision models — Body image: editorial image from The Register article on Cloudflare Clef open-weight decision models. Credit: The Register.

How Cloudflare Clef fits agent architectures

Most agency builds already split work into gather context, decide, then act. Cloudflare Clef sits in the decide step. An LLM or tool stack still gathers context and executes actions.

Gather: tickets, browser renders, chat logs, or screenshots land in state.
Decide: Clef answers urgency, owner team, severity, or allow versus deny.
Act: your Worker or backend routes, blocks, escalates, or calls a larger model.
Review: log probabilities, thresholds, and overrides for client audits.

Cloudflare also pairs Clef with Workers AI LLMs in the same network. That keeps the decision and the follow-up action close to the user when you already run on Cloudflare.

Drop-in path for Jev shops

If a client already wired Typesafe's Jev System One API, Cloudflare says Clef is compatible. Change the endpoint and model, keep the question schema, and re-run your eval suite. Details live on the Workers AI Clef model page.

Benchmarks agencies should quote carefully

Cloudflare publishes strong scores on BFCL, BANKING77, CLINC150+OOS, and home-appliance classification. On Typesafe's own workflow suite, Clef beat Jev in invoice processing, customer service, and security incidents, and trailed on agent trace observability.

The Register correctly flags that Cloudflare self-reported many Decision Index scores. For client decks, separate vendor tables from your own golden-set results. Promise a pilot, not a leaderboard win.

Hugging Face social card for the Cloudflare Clef model repository — Body image: Hugging Face social thumbnail for the Cloudflare Clef model card. Credit: Hugging Face / Cloudflare.

Local hosting and hardware reality

Open weights help regulated clients who refuse a pure SaaS path. Hardware is not casual laptop territory.

Clef-flash: about 41 GB VRAM for 64K context at single concurrency, per Cloudflare's Michelle Chen speaking to The Register.
Clef: about 85 GB VRAM under the same assumptions.
Training data: not public. Plan for open weights plus your own fine-tune data later.

For most India and USA agency pilots, start hosted on Workers AI. Move local only when a contract forces it and the GPU budget is real.

Fine-tuning and the new RL path

Cloudflare is opening a hands-on fine-tuning service with its forward-deployed engineering team. The longer plan is a self-serve loop that captures traffic, trains, and redeploys on Workers AI.

Primitives they list include AI Gateway for dataset capture, Workers AI for rollouts, Containers for RL sandboxes, a trainer for weight updates, and bring-your-own-model redeploy. That is roadmap language. Treat design-partner access as an option for clients with labeled queues, not a week-one deliverable.

Practical use cases for India and USA agency teams

Here are delivery patterns we already see in scopes. Each one maps cleanly to typed questions.

Support triage: urgent yes or no, team choice, severity score.
Trust and safety: policy scores with a human review band in the middle.
Agent guardrails: ask whether a tool call should proceed before side effects.
Threat or content labels: category choices over rendered pages or screenshots.
Visual QA: image-backed checks for storefronts, KYC docs, or UI regressions.

If you already briefed clients on AWS Strands Decider, keep the comparison honest. Read our Strands Decider 2B agency guide side by side with this Clef brief. Different vendors, same decision-model category.

Delivery checklist for Cloudflare Clef pilots

Use this checklist before you put Clef in a client production path.

Pick one workflow with a clear schema: triage, allow or deny, or severity.
Write the questions as noul, choice, or score. Cap them. Keep ids stable.
Set thresholds for auto-act, human review, and hard block.
Build a golden set of at least a few hundred labeled examples from real tickets.
Run Clef and Clef-flash against the same set. Record latency, cost, and error types.
Log everything clients will audit: state hash, answers, threshold, override reason.
Decide hosting Workers AI first unless contract language forces local GPUs.
Plan the handoff from decision to action model so ownership stays clear.

How Debugged Pro helps agencies ship this cleanly

Our software development consulting work with India and USA agencies often starts with schema design, not model shopping. We help you define the questions, thresholds, and eval set before you spend on tokens.

Browse more delivery notes on the Debugged Pro blog, or bring a live workflow and we will map it to Clef or a peer decision model with a written pilot plan.

Client briefing notes you can reuse this week

Keep the first client call short. Start with the workflow, not the model card. Explain that Cloudflare Clef answers fixed questions with probabilities your code can trust.

Then show one request and one response. Use a real ticket redacted for privacy. Walk through urgency, team, and severity. Pause on the human-review band so legal and ops see the control point.

Scope line: one workflow, one schema, one week of golden-set labeling.
Success metric: agreement with human labels plus p95 latency under the client budget.
Risk note: vendor benchmarks are directional until your set confirms them.
Exit ramp: keep the schema portable so you can swap Clef, Clef-flash, or another decision model.

India delivery pods often own overnight triage for USA clients. Edge hosting on Workers AI can shrink the round trip for those queues. USA pods usually care more about audit logs and SOC language. Log probabilities and overrides either way.

Do not oversell autonomy. Cloudflare's own examples still leave room for a human when confidence is low. That is a feature in agency contracts, not a weakness.

Key Takeaways

Cloudflare Clef and Clef-flash launched October 1, 2026 on Workers AI with Apache 2.0 weights.
They return typed probabilities for noul, choice, and score questions instead of free-form chat.
Cloudflare reports strong latency: about 209 ms median for Clef and 39 ms for Clef-flash.
Docs list hosted pricing at $0.24 and $0.09 per million input tokens for Clef and Clef-flash.
Vision, 64K context, and Jev-API compatibility make migration and multimodal triage practical.
India and USA agencies should pilot one hot-path workflow with thresholds and a golden set before scale.

About the author: Nitin Kumar is Lead Project Coordinator at Debugged Pro. He helps software and IT agencies in India and the USA plan AI delivery, agent workflows, and client-ready pilots.