Fix: Ollama reasoning models return empty content — send reasoning_effort to disable thinking

[OneByJorah]

Problem

When using Ollama as a local LLM backend with Hermes Agent, reasoning/thinking models (e.g. sorc/qwen3.5-claude-4.6-opus-q4:9b, deepseek-r1) return empty responses.

Root Cause

Ollama's OpenAI-compatible API handles reasoning models differently from other providers:

1. Reasoning models return content: (empty string) in the chat completion response
2. The actual model output is placed in a separate reasoning field
3. Hermes Agent's chat_completions transport only reads content, so it silently drops the response
4. The agent appears to hang or return blank answers

This affects all reasoning/thinking models on Ollama, including:

• sorc/qwen3.5-claude-4.6-opus-q4:9b
• deepseek-r1 (all sizes)
• qwen3.5 with thinking enabled
• Any model with thinking support in Ollama

Solution

Two changes are needed:

1. Config change (~/.hermes/config.yaml)

Add reasoning_effort: none to the agent section:

agent:
  reasoning_effort: none

2. Transport layer patch

The file agent/transports/chat_completions.py needs an Ollama-specific block in the build_kwargs function (after the LM Studio block, before extra_body assembly):

# Ollama: send reasoning_effort to disable thinking for reasoning models
# that return empty content (Ollama puts reasoning in separate field)
_base_url = params.get("base_url") or ""
_is_local = "127.0.0.1" in _base_url or "localhost" in _base_url or "::1" in _base_url
if _is_local and reasoning_config is not None and isinstance(reasoning_config, dict):
    _ollama_effort = reasoning_config.get("effort", "none") or "none"
    if _ollama_effort in {"none", "minimal", "low", "medium", "high"}:
        api_kwargs["reasoning_effort"] = _ollama_effort

This detects local Ollama endpoints and sends reasoning_effort: "none" which tells Ollama to disable the thinking/reasoning mode and return normal content responses.

Verification

After applying both changes:

• Reasoning models return proper content instead of empty strings
• The reasoning field in the response is empty (as expected with reasoning_effort: none)
• All existing non-reasoning models continue to work unchanged

Environment

• Ollama: Latest (local, http://127.0.0.1:11434/v1)
• Hermes Agent: Latest
• Hardware: RTX 3060 12GB
• Models tested: sorc/qwen3.5-claude-4.6-opus-q4:9b, granite-code:8b

References

• Ollama OpenAI compatibility docs: https://github.com/ollama/ollama/blob/main/docs/openai.md
• Related: Ollama thinking/reasoning models API behavior

[costynus]

I’d like to work on this issue

[costynus]

I tested this on current upstream/main with local Ollama 0.23.1.

Current Hermes already maps:

agent:
  reasoning_effort: none

to {"enabled": false}, and the custom / local Ollama provider already sends
extra_body["think"] = false for that path.

Relevant tests pass locally:

• scripts/run_tests.sh tests/agent/transports/test_chat_completions.py -- -q -> 81 passed
• scripts/run_tests.sh tests/run_agent/test_run_agent.py -- -q -k "ollama_think_false or ollama_no_think_param or non_custom_provider_unaffected" -> 4 passed

I also tested Ollama's OpenAI-compatible endpoint directly with the exact model
from the issue:

curl -s http://127.0.0.1:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sorc/qwen3.5-claude-4.6-opus-q4:9b",
    "messages": [
      {"role": "user", "content": "Answer in one short sentence: what is 2+2?"}
    ],
    "stream": false
  }'

On my setup, choices[0].message.content was non-empty:

"content": "2 + 2 = 4"

I also switched Hermes itself to this local Ollama model and asked 2+2. Hermes returned a visible answer (2+2 = 4.`) rather than an empty response.

I also tested:

• think: false -> content was still non-empty, but reasoning was still present
• reasoning_effort: "none" -> for this model, the separate reasoning field disappeared and content remained non-empty (2 + 2 equals 4.)

However, this does not look safe to apply broadly yet. I also tested reasoning_effort: "none" with qwen3:4b, and on that model Ollama moved the thinking text into content instead of cleanly disabling it.

So I can't reproduce the empty-content behavior on current main yet. Is there a specific Ollama version/configuration where choices[0].message.content is empty for this model?

[evlswtmn]

Confirmed this is still an issue as of Ollama latest (Jun 2026). Tested with:

• Qwen3:4b via Ollama on MacBook (192.168.1.246:11434)
• Gemma 4 12B same host
• Gateway using with OpenAI-compatible endpoint

Behavior observed:

• Qwen3 returns empty content: "" with all text in non-standard reasoning field
• Even injecting /no_think into the system prompt doesn't reliably disable thinking — model often ignores it
• Gateway logs show: finish_reason=length with prompt_tokens > max_output_tokens (thinking tokens consuming the entire output budget)
• Hermes agent sees empty response, retries, burns tokens

Also commented on Ollama upstream: ollama/ollama#15288 (comment) — the /v1 endpoint has no way to pass the think parameter that the native /api/chat supports.

Workaround that works: Switch to reasoning_effort: none or use a non-thinking model (e.g., Qwen2.5:7b instead of Qwen3, Gemma 3 instead of Gemma 4) until Ollama adds think parameter support to /v1.