Cloudflare Clef review: open-weight decision models on Workers AI

Key takeaways

• Cloudflare Clef = Clef + Clef-flash decision models on Workers AI.
• Open weights (Apache 2.0) plus managed hosting; vision / multimodal state.
• Workers AI list signal (Oct 2026): Clef ~$0.24 / M input, Clef-flash ~$0.09 / M input.
• Jev-compatible System One shape: state + up to 64 typed questions.
• Heavier hosted $/M than Jev / Laya text APIs; you pay for CF edge + multimodal + open weights path.

Table of contents

1. What Clef is
2. Pricing on Workers AI
3. Clef vs Clef-flash
4. Best for / not for
5. FAQ

---

What Clef is

Cloudflare’s Oct 1, 2026 launch positioned Clef as its answer to TypeSafe’s Jev: structured decisions, not generated prose, with open weights and edge hosting.

Cloudflare Clef turns state plus typed questions into probabilities on Workers AI. It follows the System One API family, adds multimodal inputs, and publishes Apache 2.0 weights for self-hosting.

What that means in practice:

• What it is: @cf/cloudflare/clef and @cf/cloudflare/clef-flash via REST /ai/run or env.AI.run().
• What keeps it distinct: vision, Workers AI network, open weights, optional fine-tune path.
• What it is not: a free forever chat model, and not Laya Studio’s MCP product.

Last verified: October 2026 against Cloudflare Workers AI model/pricing docs, the Clef blog, and our Cloudflare Clef tool page.

Pricing on Workers AI

Model	Input list signal	Notes
@cf/cloudflare/clef	~$0.24 / M tokens	Higher precision tier
@cf/cloudflare/clef-flash	~$0.09 / M tokens	Latency-focused tier

Billing detail that bites newcomers: Workers AI uses neurons, with a daily free neuron allocation, then Paid plan usage. Read the platform pricing page, not only the model card.

Compared with Jev’s ~$0.042/M and Laya’s ~$0.036/M text API lists, hosted Clef looks expensive per million. Cloudflare’s bet is edge latency, multimodal state, open weights, and Workers-native ops. Measure end-to-end, not sticker alone.

Clef vs Clef-flash

	Clef	Clef-flash
Size (launch)	~27B	~9B
Best for	Higher-precision decisions	Hot-path / latency-critical
Context	~64k (launch materials)	~64k
Vision	Yes (embedded images; limits apply)	Check live model card

Cloudflare publishes speed claims versus Jev (median multiples on internal runs). Treat those as vendor benchmarks. Replay your own traces.

Best for / not for

Best for:

• Cloudflare-native agents already on Workers / AI Gateway
• Teams that want open weights and a managed path
• Multimodal state (text/JSON + images) decisioning

Not ideal for:

• Pure TypeSafe SDK shops that only want first-party Jev
• Buyers optimizing solely for lowest hosted $/M text decisions
• Teams unwilling to learn Workers AI neuron billing

Related: what is a System One decision model, Jev vs Laya vs Clef.

FAQ

What is Cloudflare Clef?

Clef and Clef-flash are Cloudflare-trained System One decision models on Workers AI. They take multimodal state plus typed noul/choice/score questions and return probabilities, with a Jev-compatible API shape and Apache 2.0 open weights.

How much does Clef cost on Workers AI?

Per Cloudflare Workers AI pricing (updated around Oct 1, 2026): Clef is about $0.24 per M input tokens and Clef-flash about $0.09 per M input tokens, billed through Workers AI neurons with a daily free neuron allocation. Confirm developers.cloudflare.com pricing.

Is Clef compatible with Jev?

Cloudflare markets Clef as drop-in compatible with the System One / Jev API shape so you can change endpoint and model. Still run your own regression tests on labels and thresholds.

What is the difference between Clef and Clef-flash?

Clef (~27B) is the higher-precision model. Clef-flash (~9B) targets latency-critical hot paths. Both cite a ~64k context window in launch materials.

Can I self-host Clef?

Yes. Weights are on Hugging Face under Apache 2.0. Self-hosting needs serious GPU VRAM; many teams will use Workers AI instead.