Cloudflare’s Clef Multimodal Decision Model Now on Ollama

At a Glance
| Item | Value |
|---|---|
| Publisher | Ollama |
| Published | 2026-10-19 |
| License | MIT |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Clef, a multimodal decision model developed by Cloudflare, is now available in Ollama. It can be run in both local environments and Ollama’s cloud environment.
Clef accepts input states in the form of text or images along with typed question schemas, and directly outputs decisions and probabilities for each choice. Unlike standard large language models (LLMs) that generate tokens sequentially one by one, it is designed to evaluate each choice in the schema in parallel in a single non-autoregressive pass, featuring fast and deterministic structured outputs.
Specifications
- Parameters: 27B (Clef) / 9B (Clef-flash)
- Base Model: Qwen3.8-27B (Clef) / Qwen3.5-9B (Clef-flash)
- Context Length: 64k
- Input Modalities: Text, image
Performance
Published materials from Cloudflare include benchmark results comparing it against predecessor decision models such as Jev. The main evaluation results in the Jev Decision Index are as follows. Note that some comparison columns have been reorganized for presentation purposes.
→ Scroll horizontally to see all columns
| Benchmark | Clef | Clef-flash | Jev | DiffusionGemma Jev | Kev 9B |
|---|---|---|---|---|---|
| BFCL · case exact | 98.47 | 98.76 | 95.75 | 96.52 | 94.51 |
| ToolRet · nDCG@10 | 69.19 | 66.43 | 65.28 | 61.21 | 64.26 |
| API-Bank · accuracy | 91.93 | 93.11 | 88.19 | 83.66 | 56.30 |
| Home appliances · case exact | 82.95 | 97.73 | 52.27 | 42.05 | 25.00 |
| When2Call · accuracy | 72.37 | 65.58 | 80.97 | 75.44 | 49.62 |
| BANKING77 · macro-F1 | 94.20 | 90.93 | 79.74 | 74.28 | 84.83 |
| CLINC150+OOS · macro-F1 | 97.43 | 66.77 | 89.27 | 83.49 | 79.03 |
| BRIGHT · nDCG@10 | 45.91 | 39.26 | 47.52 | 42.94 | 38.53 |
| Amazon ESCI · macro-F1 | 57.48 | 57.39 | 55.21 | 53.37 | 49.22 |
| PhishNChips · accuracy | 79.60 | 75.05 | 62.55 | 85.35 | 50.75 |
Additionally, workflow evaluation measurements by Typesafe have been published.
| Workflow | Clef | Clef-flash | Jev |
|---|---|---|---|
| Invoice processing | 64.7 | 57.1 | 61.8 |
| Customer service | 76.3 | 77.0 | 76.0 |
| Security incidents | 62.9 | 61.7 | 61.7 |
| Agent trace observability | 68.5 | 69.8 | 71.6 |
Furthermore, inference latency (in milliseconds) measurements across 43 different benchmarks are as follows:
→ Scroll horizontally to see all columns
| Latency (ms) | Clef | Clef-flash | Jev | DiffusionGemma Jev | Kev 9B |
|---|---|---|---|---|---|
| Median | 209.3 | 38.8 | 524.1 | 84.4 | 51.4 |
| p95 | 238.6 | 122.4 | 536.0 | 211.2 | 187.9 |
Based on these measurements from the publishers, Clef recorded 98.47 (and Clef-flash 98.76) on BFCL, which measures function selection and argument assembly accuracy, showing high tool utilization accuracy that exceeds Jev’s 95.75. It also achieved higher scores than Jev in classification tasks such as BANKING77 and CLINC150+OOS, as well as in API-Bank. On the other hand, performance varies depending on the task, as it falls short of Jev’s scores in When2Call and BRIGHT, and trails DiffusionGemma Jev in PhishNChips. In terms of latency, Clef achieved a median of 209.3ms and Clef-flash 38.8ms, realizing significantly faster responses compared to Jev’s 524.1ms.
Strengths and Use Cases
Clef is a multimodal decision model equipped with a vision encoder, making it adept at directly evaluating image inputs such as screenshots, receipts, forms, and photos alongside data structures like text and JSON to make decisions. Its major strength is the ability to execute classification and judgment combining visual information in a single pass, unlike text-only decision models (such as Jev).
Model cards and official announcements specifically list the following use cases:
- Request routing: By defining destination candidates and application rules, it retrieves the selected destination and respective probabilities.
- Condition checking (judgment): By inputting Yes/No questions and evidence context, it retrieves boolean values and their probabilities.
- Policy application: Based on established rules and permitted outcomes, it receives typed decisions from provided text or images.
- Document reading: By providing page images and target evaluation fields, it extracts detailed decisions and probabilities for each field.
- Tool call moderation: Checks the validity of tool calls before AI agents execute them.
Additionally, Clef supports a 64k context window (expanded compared to Jev’s 32k), allowing more input states to be passed together. Verification by Cloudflare’s internal threat intelligence team reported that in a task capturing website screens to determine domain categories (fashion, e-commerce, phishing, etc.), a general-purpose LLM (gpt-oss-120b) returned only two classifications in 4.7 seconds, whereas Clef obtained more classification results in 2.2 seconds.
Because of its characteristic of calculating fixed typed choices and probabilities simultaneously in a single non-autoregressive pass, it is well-suited for rapid automated decision-making in agentic workflows and system control with reduced human intervention.
How to Get It
Clef (27B) and the lightweight version Clef-flash (9B) are released as open source under the Apache 2.0 license, and can be acquired and run in various environments:
- Ollama: Available for execution in local environments as well as on Ollama’s cloud.
- Hugging Face: Model weights are published, allowing local downloads for operational checks and experiments.
- Cloudflare Workers AI: Provided as a hosted API, enabling low-latency execution utilizing edge GPUs.
Because Clef maintains full API compatibility with the Jev API and System One API, existing API call code can be used as-is for a smooth replacement. Furthermore, Cloudflare provides dedicated reinforcement learning (RL) services and fine-tuning partnerships to support additional training tailored to specific business data.
Related Articles
- clef Structured Decision-Making Model: 12GB+ VRAM, GGUF Builds
- Ollama Announces Integration with Fly.io Sprites
- Ollama v0.40.4 Released with Decision Models Support
What to Read Next
- How to read BFCL → Benchmark glossary

