Ollama v0.35.1 Adds Support for Cloudflare Clef Decision Models and Expanded Search Capabilities

What changed#

Ollama released v0.35.1, introducing native support for Cloudflare’s new open-source multimodal decision models: Clef (27B) and Clef Flash (9B).

Key changes in this update include:

• Clef Decision Model Integration: Users can run clef and clef-flash using a dedicated /v1/systemone API endpoint. Requests support multimodal input, accepting base64-encoded images along with a text state parameter evaluated across structured questions to output joint confidence scores.
• Increased Web Search Depth: Models capable of performing web searches can now make up to 10 searches per response, up from the previous limit of 3.
• Explicit Modelfile Capabilities: Modelfiles now accept CAPABILITY declarations to explicitly define model functions. These declarations persist when building from safetensors or GGUF files (bringing structured metadata features alongside local execution options like Hugging Face Transformers GGUF support), during model inheritance, and upon Modelfile export.
• Capability Filtering for Clients: The ollama show command and model list now report strictly decision as the capability for decision models. This prevents client frontends from offering these models for standard chat, tool calling, or reasoning (“thinking”) tasks.
• Core Engine Updates: Underneath, Ollama updated both its llama.cpp and MLX backend engines.

Why it matters#

Decision models like Clef handle structured evaluation tasks rather than open-ended dialogue. By routing these workflows through the specialized /v1/systemone endpoint and enforcing the decision capability tag, Ollama ensures that downstream client tools do not misroute conversational or agentic tool-calling tasks to non-chat models.

Additionally, raising the web search boundary from 3 to 10 calls gives web-enabled models significantly more latitude to gather context across multi-step research queries.

Limitations#

Decision models exposed via /v1/systemone are strictly bounded to scoring state-question pairs. Because they report only the decision capability, client integrations attempting to execute standard chat completion endpoints or structured tool calls against clef or clef-flash will be rejected by compatible user interfaces.

What to do next#

To run decision queries locally on Ollama v0.35.1, pull the model and invoke the system endpoint:

1
curl http://localhost:11434/v1/systemone -d '{
2
  "model": "clef-flash",
3
  "state": "The user took this screenshot.",
4
  "images": ["<base64-encoded-image>"],
5
  "questions": {
6
    "has_ollama": {"type": "noul", "instructions": "Does this image contain Ollama?"}
7
  }
8
}'

You can verify the enforced capabilities on any local model by running ollama show <model-name> in your terminal.

Sources#

• Ollama v0.35.1 Release Notes