Ollama v0.35.1 adds the Clef decision models
Ollama v0.35.1 adds Clef and Clef Flash, Cloudflare's new open-source decision models. Both answer through the System One endpoint, and both are in the Ollama library.
Clef is 27B and Clef Flash is 9B
Two sizes, one endpoint
They come in two sizes. Clef is the larger model, at 27 billion parameters. Clef Flash is the smaller one, at 9 billion, and it's the model in the release example.
Decision requests can now include images
"model": "clef-flash", "state": "The user took this screenshot.", "images": ["<base64-encoded image>"],
Both models are multimodal. A request can send images alongside the text state, and each question is scored against both together. Images need Clef or Clef Flash with vision weights.
Clef Flash is 96 percent sure the screenshot shows Ollama
The answer is one probability
"has_ollama": {
"type": "noul",
"noul": 0.958
}
In the release example, Clef Flash is asked whether a screenshot contains Ollama. It answers with a probability of 0.958, so it is about 96 percent sure the answer is yes.
Decision models now report only the decision capability
Decision models now report only decision
In the model list and in Ollama show, a decision model now reports a single capability: decision. Apps built on Ollama stop offering these models for chat, tools or thinking.
Modelfiles can declare a CAPABILITY
Model creators can now state what a model can do with a new capability instruction in the Modelfile. It carries through GGUF and safetensors imports, inheritance, and export.
Ten web searches per response, and updated engines
Also in this release
- Web search up to ten searches per response, up from three
- Engines llama.cpp and MLX updated
- System One local server only, not yet on Ollama Cloud
Models using web search can now run up to ten searches per response, up from three. llama.cpp and the MLX engine are updated. And System One still needs a local server.
The image field is in the System One API reference
Request fields and limits
The image field, request limits and every question type are in the System One API reference, in the Ollama docs.
