Clef-Flash is a 9-billion-parameter model built to score predefined choices, not to write open-ended replies. You give it an input state and typed questions with allowed answers; it returns a probability for each option. Cloudflare announced it for Workers AI on October 1, 2026, and says its weights are available under the Apache-2.0 license.
What is Clef-Flash?
Cloudflare describes Clef-Flash as a multimodal decision model based on Qwen/Qwen3.5-9B, including its vision encoder. Its intended fit is an application that already knows what decision it needs to make—such as assigning a category, selecting an action, or rating an item against a rubric.
As an Amazon Associate I earn from qualifying purchases.
Cloudflare’s launch announcement puts the distinction plainly: “Instead of generating text, it reads an input state and a set of typed questions, then returns a probability for every allowed answer.” That makes Clef-Flash a schema-bound scorer rather than a general-purpose conversational assistant.
How does Clef-Flash work?
A request combines a state with typed questions and their permitted answers. The model scores each allowed option for each question in one forward pass; a softmax turns those scores, or logits, into per-question probabilities. Because the response is constrained to the supplied schema, the model card says the system does not generate free-form text or require output parsing.
The model card says the state can contain text, JSON, images, or video. Cloudflare’s announcement describes three question types:
noul: a yes-or-no question.choice: a question with a user-defined set of options.score: a question rated against an ordered rubric.
Cloudflare says a request can contain up to 64 questions. This structure is useful when an application needs consistent, machine-readable decisions, but it does not turn the model’s probabilities into guaranteed real-world confidence: performance still depends on the task and the choices supplied.
How is it different from a chat model?
| Aspect | Clef-Flash | General chat model |
|---|---|---|
| Input framing | A state plus typed questions and allowed answers. | Typically a conversational prompt or instruction. |
| Output | Probabilities over the permitted options; no free-form text generation, according to Cloudflare. | Generated text, which may need interpretation or parsing by an application. |
| Best-aligned use | Classification, routing, scoring, or action selection where the schema is known in advance. | Open-ended explanations, drafting, and dialogue. |
These are differences in design, not a claim that every chat model behaves identically. If your application needs a natural-language explanation or must handle unforeseen response types, Clef-Flash’s constrained output may not be enough by itself.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesHow can you run Clef-Flash?
Hosted on Workers AI
Cloudflare announced hosted access through Workers AI. Its documented model ID is @cf/cloudflare/clef-flash. Cloudflare says Clef follows the System One API, so an existing Jev integration can switch by changing its endpoint and model. That compatibility statement concerns the described integration; check the current API documentation and your own request schema before migrating.
Local or self-managed use
Cloudflare’s model card documents a local test with PyTorch 2.11 and Transformers 5.10.2 on one H200; Pillow is also needed for image and video inputs. This is the model authors’ test setup, not a stated minimum requirement or proof that a consumer GPU will be adequate. The Hugging Face page links to runtimes such as vLLM and community quantized builds, but compatibility and performance depend on the particular runtime, build, and hardware.
Cloudflare announced the model on October 1, 2026, and published its weights on Hugging Face under Apache-2.0. Cloudflare also describes hands-on fine-tuning support and says it intends to use what it learns to build a self-serve fine-tuning platform. The announcement does not establish that the self-serve platform is already available.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What do Cloudflare’s benchmarks show?
The figures below are results Cloudflare published for its 2026 evaluations, not independent replications. They use different task-specific metrics, so they should not be treated as a single universal accuracy score.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →| Evaluation | Clef-Flash | Clef | Jev | Metric and source |
|---|---|---|---|---|
| Latency across 43 benchmark runs | 38.8 ms median; 122.4 ms p95 | not stated | 524.1 ms median; 536.0 ms p95 | Latency; Cloudflare launch announcement, 2026 |
| BFCL | 98.76 | 98.47 | 95.75 | Case exact; Cloudflare launch announcement, 2026 |
| BANKING77 | 90.93 | 94.20 | 79.74 | Macro-F1; Cloudflare launch announcement, 2026 |
| CLINC150+OOS | 66.77 | 97.43 | 89.27 | Macro-F1; Cloudflare launch announcement, 2026 |
| Home appliances | 97.73 | 82.95 | 52.27 | Case exact; Cloudflare launch announcement, 2026 |
| Customer service | 77.0 | not stated | 76.0 | Exact actions; Cloudflare model card, 2026 |
| Invoice processing | 57.1 | not stated | 61.8 | Exact actions; Cloudflare model card, 2026 |
| Security incidents | 61.7 | not stated | 61.7 | Exact actions; Cloudflare model card, 2026 |
| Agent-trace observability | 69.8 | not stated | 71.6 | Primary action; Cloudflare model card, 2026 |
The pattern is mixed. Clef-Flash is ahead of Jev in Cloudflare’s latency comparison, but its benchmark results vary: it trails Clef on BANKING77 and CLINC150+OOS, and trails Jev on invoice processing and agent-trace observability. It leads the reported comparisons on BFCL and home appliances. Cloudflare positions the 9B model for latency-sensitive decisions and the 27B Clef for highest-precision decisions; the task-level results are more informative than that broad positioning when choosing between them.
What is established about “I spotted it on DEV·TV”?
The official Cloudflare materials reviewed here explain Clef and Clef-Flash, but do not explain DEV·TV or verify the first-person discovery context in the original title. The article therefore treats the product’s documented design and launch as the subject, without speculating about where or how it was first spotted.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
