Clef stops writing answers word by word: one pass scores every question and returns a probability
Synopsis
Cloudflare's 27B multimodal decision model Clef is now available on Ollama, running locally or in Ollama's cloud. It takes a state plus a schema of typed questions and, in one pass over text, JSON or images, scores every answer to every question at once instead of generating a response word by word. On Cloudflare's published Jev Decision Index evaluations Clef beats Jev on most benchmarks, but trails it on When2Call and BRIGHT.
Interpretation
Decision tasks get a dedicated model entry point: given a state and a schema of typed questions, it returns the decision itself and a probability for each candidate answer rather than free text. The prior pattern had a general model write a response word by word, with external logic parsing it into a decision; Clef replaces that step with direct scoring per question. On Cloudflare's published Jev Decision Index evaluations, Clef scores 98.47 case exact on BFCL and 91.93 accuracy on API-Bank, against Jev's 95.75 and 88.19; on When2Call Clef reaches 72.37, below Jev's 80.97.
The model accepts text, JSON and images, covering screenshots, receipts, forms and photos. Multimodal input and structured decisions sit in one model instead of one model reading the image and another module making the call. Clef is fine-tuned from Qwen3.8-27B, released under the Apache 2.0 license, and compatible with the Jev and System One APIs.
The same scoring approach covers routing a request, checking a condition, applying a policy, reading document fields and moderating an agent's tool call. These tasks previously tended to need separate rules or a classifier each. On Typesafe's own workflow evals, Clef beats Jev in three of four areas, scoring 68.5 on agent trace observability against Jev's 71.6.
Perspective
The report covers the release of Clef on Ollama and the benchmark and workflow results Cloudflare published. It is aimed at engineers who need request routing, condition checks, policy application, document field reading or agent tool-call moderation expressed as decisions with probabilities. The model accepts text, JSON and image input, runs locally or in Ollama's cloud, and is released under the Apache 2.0 license.
Clef does not lead on every evaluation: When2Call accuracy is 72.37 against Jev's 80.97, BRIGHT nDCG@10 is 45.91 against Jev's 47.52, and PhishNChips accuracy is 79.60 against DiffusionGemma Jev's 85.35. On Typesafe's workflow evals it also trails Jev on agent trace observability. On latency, Clef's median is 209.3 ms, above Clef-flash's 38.8 ms.