Cloudflare's Clef decision model now runs locally and on Ollama's cloud
Clef, a 27B multimodal decision model from Cloudflare, is now available on Ollama, where users can run it locally or on Ollama's cloud. According to Ollama, the model turns a state and a schema of typed questions into decisions, and it accepts both text and image input.
- Clef, a 27B multimodal decision model from Cloudflare, is now on Ollama, locally or in the cloud.
- It is fine-tuned from Qwen3.8-27B and released under the Apache 2.0 license.
- According to Ollama, it scores every answer to every question in one pass instead of generating text word by word.
- Clef beats Jev on most of the benchmark suite, but trails it on tasks such as When2Call and BRIGHT.
- Median latency across 43 eval benchmarks is 209.3 ms for Clef, against 524.1 ms for Jev.
What Clef does and how it is built
Clef is fine-tuned from Qwen3.8-27B and released under the Apache 2.0 license. Ollama says it is compatible with the Jev and System One APIs. It can make decisions over text, JSON or images, including screenshots, receipts, forms and photos.
The model works differently from a typical chat model. According to Ollama, Clef reads the input once and scores every answer to every question at the same time, instead of writing a response word by word. The result is a typed decision with probabilities rather than free-form text.
Use cases listed by Ollama
Ollama names five tasks Clef is meant for:
- Route a request: define destinations and when each applies. Clef returns the chosen destination and each destination's probability.
- Check a condition: define a yes-or-no question and the evidence. Clef returns true or false and the probability of each.
- Apply a policy: define rules and allowed outcomes to receive a typed decision based on supplied text or images.
- Read a document: supply a photo of a page and the fields to evaluate, and receive field-by-field decisions with probabilities.
- Tool-call moderation: check an agent's tool call before it runs.
Benchmarks and latency
The figures come from Cloudflare's announcement and use evaluations from the Jev Decision Index. Ollama says Clef tops Cloudflare's Decision Index and beats Jev on most of its benchmark suite.
Clef scores 98.47 on BFCL (case exact) against 95.75 for Jev, and 91.93 on API-Bank (accuracy) against 88.19. On Home appliances (case exact) it reaches 82.95 against 52.27, and on BANKING77 (macro-F1) 94.20 against 79.74. On CLINC150+OOS (macro-F1) it scores 97.43 against 89.27. Jev is ahead on When2Call (80.97 versus 72.37) and BRIGHT (47.52 versus 45.91). On PhishNChips, DiffusionGemma leads with 85.35 against 79.60 for Clef.
On Typesafe's own workflow evals, Clef beats Jev in 3 of 4 areas. It scores 64.7 on invoice processing (Jev 61.8), 76.3 on customer service (Jev 76.0) and 62.9 on security incidents (Jev 61.7). Jev leads on agent trace observability, 71.6 to 68.5.
Across 43 eval benchmarks, Clef has a median latency of 209.3 milliseconds and a p95 of 238.6 milliseconds. Jev is listed at 524.1 and 536.0 milliseconds. The smaller Clef-flash variant is listed at a median of 38.8 milliseconds and a p95 of 122.4 milliseconds.
What it means for developers
For developers who run models through Ollama, Clef adds an open, Apache 2.0 licensed option for structured decisions such as routing, policy checks and document reading, either on their own hardware or in Ollama's cloud. Compatibility with the Jev and System One APIs, as stated by Ollama, may ease integration into existing setups. The benchmark results are reported by Cloudflare and Ollama, and the tables show that the lead over other models varies by task.
Ollama points to the Clef model page and its Ollama Cloud documentation for further details.
Ollama: Clef is available on Ollama's cloud
