Skip to content

Example Research Reports

Explore research reports generated by MAESTRO using various Large Language Models, showcasing quality, style, and capabilities across different model sizes and deployment options.

Overview

These example reports demonstrate:

  • Model Capabilities - Quality differences between model sizes
  • Writing Styles - Various tones and formats achieved through prompting
  • Research Depth - How models handle complex topics
  • Context Handling - Long-form content generation abilities
  • Deployment Options - Self-hosted vs cloud-based models

Note on Report Quality

These reports include examples of mistakes from various models. We are improving the framework to try and catch some of these mistakes, but due to the unpredictable nature of these, it is impossible to account for all eventualities. It is impossible to design a system that will work well with all models. If the model you're using keeps producing bad results, you may want to try a different model.

Cloud-Based Model Combinations

Reports generated using API-based models (OpenAI, Google Gemini) in various combinations for Fast, Mid, and Intelligent tiers.

Self-Hosted Open Source Models

Reports generated using locally deployed models via vLLM/SGLang, providing complete data privacy and customization.

Large Models (70B+ Parameters)

  • Qwen 2.5 72B

    One of the most capable open models, excellent for complex research.

    Highlights:

    • Superior reasoning and analysis
    • Excellent context handling
    • Strong technical writing

    View Reports

  • GPT-OSS 120B

    The largest open model tested, exceptional comprehension.

    Highlights:

    • Best-in-class synthesis
    • Creative writing excellence
    • Deep technical analysis

    View Reports

Medium Models (20-34B Parameters)

  • Qwen 3 32B

    Excellent balance of quality and performance.

    Highlights:

    • Fast inference speed
    • Strong general capabilities
    • Good context retention

    View Reports

  • Qwen 3 30B-A3B

    Specialized variant with unique strengths.

    Highlights:

    • Strong economic analysis
    • Creative content generation
    • Policy research expertise

    View Reports

  • Gemma 3 27B

    Google's model with domain-specific strengths.

    Highlights:

    • Academic writing
    • Business strategy
    • Professional tone

    View Reports

  • GPT-OSS 20B

    Smaller but efficient for focused tasks.

    Highlights:

    • Fast inference
    • Good for summaries
    • Psychology/business topics

    View Reports

Research Styles

Explore reports organized by writing style and target audience.

  • Academic & Technical

    Scholarly papers with citations, methodology sections, and technical depth.

    Best for: Research papers, technical documentation, academic analysis

  • Popular Science

    Accessible explanations of complex topics for general audiences.

    Best for: Blog posts, magazine articles, educational content

  • Business & Professional

    Executive summaries, market analyses, and strategic recommendations.

    Best for: Business reports, market research, strategy documents

  • Creative & Narrative

    Storytelling approaches including travel writing and fictional narratives.

    Best for: Travel guides, creative content, narrative reports

Performance Observations

Local Model Insights

Context Usage Patterns

The above reports were generated using the following context settings in the Settings → Research Tab:

Context Settings Tested:

  • Planning: 100,000 - 150,000 tokens
  • Writing preview: 20,000 - 30,000 tokens
  • Note content: 10,000 - 25,000 tokens
  • Writing agent: 200,000 - 250,000 tokens

KV Cache Observations:

  • Normal usage: ~25% with 25 concurrent requests
  • Peak usage: 80% of 100K context observed (multiple reports running at the same time)
  • Extreme case: 97% with 12 concurrent requests (multiple reports running at the same time)
  • Most users unlikely to reach limits during normal operation

Model-Specific Issues

  • Qwen 3 30B-A3B: Issues generating valid structured outputs
  • GPT-OSS 20B: Instruction following problems, especially in non-traditional reports

Cloud vs Local Comparison

Aspect Cloud Models Local Models
Privacy Data sent to provider Complete data control
Cost Per-token pricing One-time hardware cost
Latency Faster in most cases Speed depends on hardware
Availability Subject to rate limits Always available
Model Selection Proprietary & Open Weights models Any open model
Maintenance Provider managed Self-managed

Getting Started

  1. Choose Deployment Type:

    • Cloud: Quick setup, no hardware required
    • Local: Privacy-focused, requires GPU

  2. Select Model Tier:

    • Fast: Quick responses, simple tasks
    • Mid: Balanced performance
    • Intelligent: Complex analysis

  3. Configure in Settings:

    • Navigate to Settings → AI Config
    • Select provider and models
    • Test configuration

  4. Start Researching:

    • Use example prompts as inspiration
    • Adjust parameters for your needs
    • Experiment with different models