REASONING_MODELS.md

Reasoning Models in MuseWeb

MuseWeb automatically detects and handles models with reasoning/thinking capabilities. For web page generation, thinking output is automatically disabled to ensure clean HTML responses.

How It Works

When MuseWeb detects a reasoning model (based on configurable patterns), it automatically:

1. Sends thinking: false in the API request to disable reasoning output
2. Ensures clean responses without thinking tags cluttering the web page
3. Maintains full functionality while providing a better user experience

Supported Models

The following model patterns are automatically detected as reasoning models:

DeepSeek Models

• deepseek-r1-distill - DeepSeek R1 distilled models (most specific)
• deepseek - All DeepSeek models (general)

OpenAI R1 Models

• r1-1776 - OpenAI R1-1776 models

Qwen Models

• qwen3 - Qwen3 models (specific)
• qwen - All Qwen models (general)

Perplexity Models

• sonar-reasoning-pro - Perplexity Sonar reasoning pro models
• sonar-reasoning - Perplexity Sonar reasoning models

Gemini Models

• gemini-2.5-flash-lite-preview-06-17 - Specific Gemini model
• gemini-2.5-flash - Gemini 2.5 Flash models

Configuration

You can customize which models are treated as reasoning models by editing the reasoning_models section in your config.yaml:

model:
  reasoning_models:
    - "deepseek-r1-distill"  # Most specific patterns first
    - "sonar-reasoning-pro"
    - "sonar-reasoning"
    - "r1-1776"
    - "qwen3"
    - "deepseek"             # General patterns last
    - "qwen"

Pattern Matching Rules

1. Case-insensitive: Patterns match regardless of case
2. Substring matching: Pattern must appear anywhere in the model name
3. Priority-based: First match wins, so list specific patterns before general ones
4. Automatic thinking disabled: All matched models get thinking: false automatically

Why Thinking is Disabled

For web page generation, reasoning output creates several problems:

• Cluttered HTML: Thinking tags appear in the generated web page
• Poor UX: Users see the model's internal reasoning process
• Inconsistent output: Some responses have thinking, others don't
• Parsing issues: Thinking tags can interfere with HTML structure

By automatically disabling thinking for reasoning models, MuseWeb ensures:

• Clean web pages without reasoning clutter
• Consistent output across all requests
• Better user experience with focused content
• Proper HTML structure without interference

Adding New Models

To add support for new reasoning models, simply add their name patterns to the reasoning_models list in your config:

model:
  reasoning_models:
    - "new-reasoning-model"  # Add your pattern here
    - "another-pattern"

No code changes are required - the system will automatically detect and handle new patterns.