Skip to main content

🧠 Model Choice

ToothFairyAI provides a comprehensive selection of AI models optimised for different use cases and complexity levels. We embrace the open-source community by supporting all major open-source models alongside our proprietary offerings, ensuring you have access to the best of both worlds. Each model is designed to excel in specific scenarios, from lightweight assistant tasks to complex reasoning and code generation, with a commitment to keeping pace with the latest open-source innovations.

🧚 ToothFairyAI Models (Sorcerer and Mystica)​

What Are Sorcerer & Mystica?​

Think of these models as AI orchestrators rather than single models. Behind the scenes, Sorcerer and Mystica operate a sophisticated routing system that manages a diverse pool of models - like having a team of specialists instead of one generalist. This proprietary architecture combines carefully selected open-source models with ToothFairyAI's custom fine-tuned models, creating a unique ensemble that dynamically picks the perfect model for each specific task.

Core Architecture:

  • Dynamic Model Routing: Real-time intelligent selection from a curated pool based on task requirements
  • Hybrid Model Pool: Combines open-source foundations with proprietary fine-tuning
  • Advanced Capabilities: Enables features that surpass state-of-the-art closed-source models in specific domains, including dynamic graph generation and multi-modal RAG with contextual image retrieval
  • Thinking Variants: Both series offer enhanced reasoning versions for complex problem-solving

Agentic Specialisation: The Mystica models are specifically optimised for agentic workflows, excelling in:

  • Superior Tool Integration: Advanced function calling and API orchestration through specialised routing
  • Multi-step Planning: Optimised for complex task decomposition and execution
  • Context Management: Efficient handling of long-running agent conversations and state
  • Adaptive Reasoning: Automatically adjusts reasoning depth based on task complexity

Advanced Model Architecture​

🧚 Sorcerer and Mystica represent a breakthrough in AI model deployment through their sophisticated routing architecture. Unlike traditional single-model approaches, these systems employ an intelligent router that manages a diverse pool of models, each optimised for specific tasks.

Technical Innovation​

The routing system analyses incoming requests in real-time, considering factors such as task complexity, required capabilities, and response constraints. It then dynamically selects the optimal model or combination of models from the pool. This approach enables:

  • Specialised Performance: Each model in the pool can be highly optimised for specific tasks rather than being a generalist
  • Adaptive Scaling: The system automatically scales computational resources based on task requirements
  • Continuous Improvement: New models can be added to the pool without disrupting existing workflows
  • Cost Optimisation: Efficient routing ensures that only the necessary computational resources are used
Enterprise Deployment​

For enterprise customers, ToothFairyAI offers the ability to:

  • Host Locally: Deploy the entire routing system and model pool on-premises
  • Customise Model Selection: Adjust routing preferences and model weights based on specific business needs
  • Add Custom Models: Integrate proprietary or specialised models into the existing pool
  • Fine-tune Components: Modify individual models within the pool for domain-specific performance

This architecture enables capabilities that are difficult to achieve with monolithic models, such as seamless integration of dynamic visualisations, contextual multi-modal retrieval, and adaptive reasoning depth based on task complexity.

By selecting the appropriate model for your specific needs, you can optimise both performance and efficiency in your ToothFairyAI agent implementations.

Understanding Model Categories​

Our models are organised into three main categories based on computational power and capabilities:

  • Large Models: High-performance models for complex reasoning, advanced coding, and demanding tasks
  • Medium Models: Balanced performance models suitable for most business applications
  • Small Models: Efficient models optimised for speed and cost-effectiveness

Model Selection by Use Case​

Code Generation & Reasoning​

For Complex Development Tasks:

  • Deepseek R1 05/28 (Large) - Advanced reasoning capabilities for complex problem-solving
  • Deepseek V3 03-24 (Large) - High-performance code generation and architectural design

For Standard Development:

  • Qwen2.5-32B-Coder (Medium) - Reliable code generation with good performance balance

Code Completion & Bug Fixing​

For Production Code:

  • Qwen2.5-32B-Coder (Medium) - Excellent at understanding context and fixing issues

General Reasoning & Planning​

For Smart Assistant Applications:

  • Sorcerer 1.5 Thinking (Medium) - Optimised for low-medium effort reasoning tasks with enhanced thinking capabilities
  • DeepSeek R1 05/28 (Large) - Superior analytical thinking for complex planning
  • Deepseek V3 03-24 (Large) - Advanced strategic planning and decision-making

For Business Applications:

  • Qwen2.5-72B-Instruct (Medium) - Strong reasoning for business logic
  • Llama 3.3 70B (Medium) - Reliable planning and analysis capabilities

Function Calling & Tool Use​

For Agentic Workflows:

  • Mystica (Medium) - Specialised for low thinking agentic tasks with excellent tool integration
  • Mystica 1.5 Thinking (Medium) - Enhanced reasoning for the hardest agentic problems
  • Qwen3 235B A22B (Large) - Maximum capability for complex tool orchestration

For Standard Integration:

  • Qwen 3 Family Models (Large/Medium/Small) - Consistent function calling across different scales

Long Context & Summarization​

For Document Processing:

  • Llama 4 Maverick (Large) - Exceptional long-context understanding for comprehensive analysis

For Standard Tasks:

  • Llama 4 Scout (Medium) - Efficient summarisation with good context retention

Vision & Document Understanding​

For Advanced Visual Tasks:

  • Llama 4 Maverick (Large) - Superior image analysis and document comprehension

For Standard Visual Processing:

  • Qwen2.5-VL 32B Instruct (Medium) - Reliable visual understanding
  • Qwen2.5-VL 72B Instruct (Medium) - Enhanced visual reasoning
  • Llama 4 Scout (Medium) - Balanced visual processing capabilities

For Quick Visual Tasks:

  • Qwen2.5-VL 3-7B (Small) - Fast image processing with reduced computational requirements

Low-Latency NLU & Extraction​

For Real-Time Applications:

  • Llama 3.1 8B (Small) - Ultra-fast natural language understanding
  • Llama 3.2 3B (Small) - Efficient entity extraction
  • Llama 3.2 1B (Small) - Minimal latency for simple extraction tasks
  • Qwen3 8B (Small) - Quick processing with good accuracy

Deepseek Series​

  • Code Excellence: Industry-leading code generation and understanding
  • Advanced Reasoning: Superior analytical thinking for complex problems
  • Latest Architecture: Cutting-edge model design for optimal performance

Model Recommendations by Agent Type​

Assistant Agents​

Recommended Models:

  • Sorcerer 1.5 Thinking (Medium) - Ideal for everyday smart assistance with enhanced reasoning
  • Mystica (Medium) - Excellent for general-purpose interactions requiring minimal thinking overhead

Operator Agents​

Recommended Models:

  • Mystica (Medium) - Perfect for complex retrieval and analysis tasks
  • Kimi K2 (Large) - Superior for document-heavy knowledge work
  • DeepSeek R1 05/28 (Large) - Excellent analytical reasoning for research tasks

Programmer Agents​

Recommended Models:

  • Deepseek R1 05/28 (Large) - Advanced code reasoning and architecture
  • Qwen2.5-32B-Coder (Medium) - Balanced performance for most coding tasks
  • Qwen3 14B (Small) - Quick coding assistance and simple tasks

Orchestrator Agents​

Recommended Models:

  • Mystica 1.5 Thinking (Medium) - Optimal for complex multi-step task planning
  • Qwen3 235B A22B (Large) - Maximum capability for intricate workflow orchestration
  • DeepSeek R1 05/28 (Large) - Superior reasoning for task breakdown and delegation

Performance Considerations​

When selecting a model, consider these factors:

  1. Task Complexity: Match model size to task requirements
  2. Response Speed: Smaller models provide faster responses
  3. Cost Efficiency: Balance performance needs with resource usage
  4. Use Case Specificity: Choose models optimised for your primary use case

Best Practices​

  1. Start Medium: Begin with medium-sized models for most applications
  2. Scale as Needed: Upgrade to larger models when complexity demands it
  3. Test Performance: Evaluate different models with your specific use cases
  4. Monitor Usage: Track model performance and adjust selections accordingly
  5. Consider Context: Factor in the agent type and intended workflow when choosing

Why Aren't Sorcerer & Mystica Open Source?​

We love open source - we really do! That's why we support every major open-source model out there. But here's the thing about Sorcerer and Mystica: they're not just models, they're entire orchestration systems with years of R&D, custom routing logic, and specially fine-tuned models that we've trained on proprietary datasets.

Think of it this way: we're like a restaurant that uses open-source recipes (all the amazing open models) but has also developed our own secret sauce (the routing system and fine-tuned models). We share the kitchen with you, let you bring your own recipes, and even let you run the whole restaurant on your premises - but the secret sauce recipe stays in the vault.

Why this approach works for everyone:

  • You get access to ALL open-source models plus our specialized systems
  • We can keep investing in making the magic happen (R&D isn't free!)
  • Enterprises get production-grade support and guarantees
  • The open-source community benefits from our infrastructure and integration work

So yes, Sorcerer and Mystica's secret ingredients stay secret - but everything else is as open as it gets!