Clef vs Clef-Flash vs Clef-omni
Cloudflare publishes three Clef decision models. They share question types (noul, choice, score) and return probabilities. They differ in size, price, hosted context and whether they natively read audio and video.
URL kept as /clef-vs-clef-flash for existing search rankings. Independent site — not Cloudflare official docs.
Compare Flash, Omni & Clef in the live playground → Or rate & sort photos in bulk →
| Clef-Flash | Clef-omni | Clef | |
|---|---|---|---|
| Parameters | 9B | 30B MoE (~3B active) | 27B |
| Price (input) | $0.038 / M | $0.15 / M | $0.24 / M |
| Hosted context | 24,000 | 64,000 | 64,000 |
| Inputs | text, JSON, images | text, JSON, images, audio, video | text, JSON, images |
| Question types | noul, choice, score | noul, choice, score | noul, choice, score |
| Model id | @cf/cloudflare/clef-flash | @cf/cloudflare/clef-omni | @cf/cloudflare/clef |
When to use Clef-Flash
Default for high-volume text and image decisions: routing, tagging, photo sorting. Cheapest hosted option; this site’s sorter and playground stay on Flash.
When to use Clef-omni
Native audio and video in one call — call QA, install acceptance, short-video policy checks. Priced below Clef. Product page: Clef-omni.
When to use Clef
When Flash probabilities are close or wrong on your examples, or you need the larger 64k text context. Common pattern: Flash first, escalate low-confidence cases to Clef.
Published benchmarks (Clef vs Flash)
Figures Cloudflare and Typesafe published (as listed on Ollama’s model pages). We didn’t run them. Higher is better; the better model in each row is bold. Omni has separate official tables on the 2026-10-09 blog.
| Benchmark | Clef | Clef-Flash |
|---|---|---|
| BANKING77 (macro-F1) | 94.2 | 90.9 |
| CLINC150+OOS (macro-F1) | 97.4 | 66.8 |
| ANLI (macro-F1) | 69.8 | 59.1 |
| When2Call (accuracy) | 72.4 | 65.6 |
| PhishNChips (accuracy) | 79.6 | 75.0 |
| BFCL (case exact accuracy) | 98.5 | 98.8 |
| API-Bank (accuracy) | 91.9 | 93.1 |
| Home appliance simulator (case exact accuracy) | 83.0 | 97.7 |
| RouterBench (selected quality) | 79.7 | 79.9 |
Business workflows (Typesafe Evals)
| Workflow | Clef | Clef-Flash |
|---|---|---|
| Invoice processing | 64.7 | 57.1 |
| Customer service | 76.3 | 77.0 |
| Security incidents | 62.9 | 61.7 |
| Agent trace observability | 68.5 | 69.8 |
Latency (Flash vs Clef, published)
Median 38.8 ms for Clef-Flash vs 209.3 ms for Clef. Official Omni medians (blog): text ~130 ms, image ~150 ms, 21 s video with sound ~1.5 s — not measured by us.
Compare them yourself
All three models are in the playground. For audio/video pick Clef-omni (upload or load a sample). Production volume: API example.
FAQ
Is Clef more accurate than Clef-Flash?
On most published text benchmarks, yes: for example 97.4 vs 66.8 macro-F1 on CLINC150+OOS. Clef-Flash wins a few and is much faster and cheaper. Your own data is still the real test.
Do they use the same API?
Yes for questions and answers. Change the model id in the URL and the model field. Only Clef-omni accepts audio and videos fields.
Which one is cheaper?
Clef-Flash at $0.038/M input is cheapest. Clef-omni at $0.15/M is cheaper than Clef ($0.24/M) and is the only native audio/video option.
When should I use Clef-omni?
When the evidence is audio or video (call recordings, install clips, short-form moderation). For text and still images at high volume, stay on Flash; escalate hard cases to Clef.
Model facts from Workers AI docs and the Cloudflare blog, checked 2026-10-11. Flash/Clef benches from Ollama model pages, checked 2026-10-03.