openclaw - 💡(How to fix) Fix [Bug]: Ollama cloud models (gemini-3-flash-preview:cloud, deepseek-v4-pro:cloud) fail with "Assistant turn failed before producing content" or silent fallback [2 comments, 3 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Two Ollama cloud models that work perfectly via direct Ollama API call fail when used through OpenClaw:

1. ollama/gemini-3-flash-preview:cloud → "Assistant turn failed before producing content"
2. ollama/depseek-v4-pro:cloud → Silent fallback to next model in chain Both models work fine when called directly via curl http://localhost:11434/api/generate.

Steps to reproduce

1. Add ollama/gemini-3-flash-preview:cloud as primary or fallback model
2. Send any message to trigger the model
3. Observe error: "Assistant turn failed before producing content"
4. OR for deepseek: Observe silent fallback to next model (kimi)

Expected behavior

Models should respond through OpenClaw just as they do via direct Ollama API calls.

Actual behavior

• gemini-3-flash-preview:cloud: Error message "Assistant turn failed before producing content"
• deepseek-v4-pro:cloud: Model silently falls back to next in chain (kimi-k2.6:cloud)

OpenClaw version

2026.4.26 (be8c246)

Operating system

OS: Linux 6.17.13-4-pve (Proxmox VM)

Install method

Node: v22.22.2 npm global

Model

Models affected: ollama/gemini-3-flash-preview:cloud ollama/depseek-v4-pro:cloud Working models (for comparison): ollama/kimi-k2.6:cloud ✅ ollama/qwen3.5:cloud ✅ ollama/glm-5.1:cloud ✅

Provider / routing chain

openclaw -> local ollama server -> ollama.com -> models

Additional provider/model setup details

Ollama version: Latest (deepseek-v4-pro:cloud pulled 2026-04-28) Models affected: ollama/gemini-3-flash-preview:cloud ollama/depseek-v4-pro:cloud Working models (for comparison): ollama/kimi-k2.6:cloud ✅ ollama/qwen3.5:cloud ✅ ollama/glm-5.1:cloud ✅

Logs, screenshots, and evidence

Direct API verification
Copy
# Works fine - 24s response time
curl http://localhost:11434/api/generate -d '{
  "model": "deepseek-v4-pro:cloud",
  "prompt": "Write a Python hello world",
  "stream": false
}'
Configuration
Copy
{
  "model": {
    "fallbacks": [
      "ollama/qwen3.5:cloud",
      "ollama/glm-5.1:cloud",
      "ollama/depseek-v4-pro:cloud",
      "ollama/gemma4:31b-cloud",
      "ollama/gemini-3-flash-preview:cloud"
    ],
    "primary": "ollama/kimi-k2.6:cloud"
  }
}
Possible related issues
#64710 — Similar timeout issue with Ollama models (fixed in 2026.4.26 for local models, but cloud models may have different capability discovery)
#57626 — Model fallback cascades
Additional notes
Both models work perfectly in other software (e.g., Frigate for gemini)
The fallback chain correctly routes to kimi when these models fail
This appears to be an OpenClaw-specific compatibility issue, not an Ollama issue

Impact and severity

No response

Additional information

No response