mindXdashboard/docs/book/journal/api/dojo/inference/governance/origin
philosophymanifestothesisoriginwhitepaperataraxiaroadmappress|archovervieworchestrationcodebasehierarchycore|agentsmindXagentceomastermindbdievolutionauthorall
govdaiocivilizationidentitysecurity|memorypgvectorembedaglmmemory|inferencevllmollamamistralgemini|timeoracle
toolsindextoolsa2amcpshell|publishauthoragentbookjournal|deployproductionsecuritymonitoring|apireferenceswagger|learnusageguidehackathon
mistral_models.md · 8.5 KB

Mistral AI Models Configuration

Overview

The models/mistral.yaml file provides comprehensive configuration for all Mistral AI models available through their API. This configuration follows the same structure as the Gemini models configuration, ensuring consistency across the MindX system.

Model Categories

Large Language Models (LLMs)

Mistral Large Models

  • mistral-large-latest - Latest large model with excellent reasoning and writing capabilities
  • mistral-large-2402 - Specific version from February 2024
  • mistral-8x22b-instruct - Mixture of experts model with 22B parameters per expert

    • Mistral Small Models

  • mistral-small-latest - Latest small model optimized for speed and efficiency
  • mistral-small-2402 - Specific version from February 2024
  • mistral-nemo-latest - Ultra-fast model for high-throughput applications
  • mistral-nemo-12b-latest - Larger Nemo variant with 12B parameters

    • Code Generation Models

  • codestral-latest - Latest code generation model
  • codestral-22b-latest - Larger code model with 22B parameters
  • codestral-2405 - Specific version from May 2024

    • Legacy Models

  • mistral-7b-instruct - Original 7B instruction-tuned model
  • mistral-7b-instruct-v0.3 - Version 0.3 of the 7B model
  • mistral-8x7b-instruct - Mixture of experts with 7B parameters per expert
  • mistral-8x7b-instruct-v0.1 - Version 0.1 of the 8x7B model

    • Embedding Models

  • mistral-embed - Original embedding model
  • mistral-embed-v2 - Improved embedding model with better performance

    • Configuration Structure

    Each model entry includes the following parameters:

    Task Scores (0.0 - 1.0)

  • reasoning - Logical reasoning and problem-solving ability
  • code_generation - Code writing and programming capabilities
  • writing - Creative and technical writing quality
  • simple_chat - Conversational and chat capabilities
  • data_analysis - Data processing and analysis skills
  • speed_sensitive - Performance in time-critical applications

    • Pricing (per 1K tokens)

  • cost_per_kilo_input_tokens - Cost for input tokens
  • cost_per_kilo_output_tokens - Cost for output tokens

    • Technical Specifications

  • max_context_length - Maximum context window size
  • supports_streaming - Whether the model supports streaming responses
  • supports_function_calling - Whether the model supports function calling
  • api_name - The actual API identifier for the model

    • Capabilities

  • text - General text processing
  • reasoning - Advanced reasoning capabilities
  • code_generation - Code writing and programming
  • multilingual - Multi-language support
  • fill_in_middle - Fill-in-the-middle code completion
  • embedding - Text embedding generation

    • Model Selection Guidelines

    For General Use

  • Best Overall: mistral-large-latest - Excellent balance of capabilities
  • Fast & Efficient: mistral-small-latest - Good performance with lower cost
  • Ultra-Fast: mistral-nemo-latest - Maximum speed for high-throughput needs

    • For Code Generation

  • Best Code Model: codestral-latest - Specialized for programming tasks
  • Advanced Code: codestral-22b-latest - More powerful for complex code
  • Fill-in-Middle: codestral- models support FIM for code completion

    • For Embeddings

  • Latest: mistral-embed-v2 - Improved performance and quality
  • Legacy: mistral-embed - Original embedding model

    • For Cost Optimization

  • Budget Option: mistral-nemo-latest - Lowest cost per token
  • Balanced: mistral-small-latest - Good performance-to-cost ratio
  • Premium: mistral-large-latest - Best quality for complex tasks

    • Context Length Considerations

    128K Context Models

  • mistral-large-latest
  • mistral-small-latest
  • mistral-nemo-latest
  • codestral-latest
  • codestral-22b-latest
  • mistral-nemo-12b-latest

    • 64K Context Models

  • mistral-8x22b-instruct

    • 32K Context Models

  • mistral-7b-instruct
  • mistral-8x7b-instruct
  • mistral-7b-instruct-v0.3
  • mistral-8x7b-instruct-v0.1

    • 8K Context Models

  • mistral-embed
  • mistral-embed-v2

    • Performance Characteristics

    Speed vs Quality Trade-offs

    ModelSpeedQualityCostBest For mistral-nemo-latest⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐High-throughput, simple tasks mistral-small-latest⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐Balanced performance mistral-large-latest⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐Complex reasoning, quality codestral-latest⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐Code generation

    Task-Specific Recommendations

    Reasoning Tasks

  • Primary: mistral-large-latest (0.92 score)
  • Alternative: mistral-8x22b-instruct (0.90 score)

    • Code Generation

  • Primary: codestral-22b-latest (0.96 score)
  • Alternative: codestral-latest (0.95 score)

    • Writing Tasks

  • Primary: mistral-large-latest (0.94 score)
  • Alternative: mistral-8x22b-instruct (0.93 score)

    • Chat/Conversation

  • Primary: mistral-large-latest (0.96 score)
  • Alternative: mistral-small-latest (0.94 score)

    • Data Analysis

  • Primary: mistral-8x22b-instruct (0.91 score)
  • Alternative: mistral-large-latest (0.89 score)

    • Integration with MindX

    Default Model Selection

    The MindX system uses these models based on configuration:

    # In .env file
MINDX_LLM__MISTRAL__DEFAULT_MODEL="mistral-large-latest"
MINDX_LLM__MISTRAL__DEFAULT_MODEL_FOR_CODING="codestral-latest"
MINDX_LLM__MISTRAL__DEFAULT_MODEL_FOR_REASONING="mistral-large-latest"

    Model Registry Integration

    The models are automatically registered in the MindX model registry and can be selected based on:

  • Task requirements - Automatic selection based on task scores
  • Cost constraints - Selection based on pricing tiers
  • Performance needs - Selection based on speed requirements
  • Context length - Selection based on input size

    • Graceful Degradation

    If Mistral API keys are not available, the system will:
  • Log warnings about missing API keys
  • Fall back to other LLM providers (Ollama, Gemini)
  • Provide mock responses for testing

    • Usage Examples

    Basic Model Selection

    from llm.llm_factory import LLMFactory
    

    Get a Mistral model
    
model = await LLMFactory.create_llm_handler(
    provider="mistral",
    model="mistral-large-latest"
)

    Code Generation

    # Use Codestral for code generation
code_model = await LLMFactory.create_llm_handler(
    provider="mistral",
    model="codestral-latest"
)

    High-Speed Processing

    # Use Nemo for high-throughput tasks
fast_model = await LLMFactory.create_llm_handler(
    provider="mistral",
    model="mistral-nemo-latest"
)

    Monitoring and Analytics

    Performance Metrics

    The system tracks:
  • Token usage - Input/output token consumption
  • Response times - Model latency and throughput
  • Cost tracking - Real-time cost monitoring
  • Quality scores - Task-specific performance metrics

    • Model Comparison

    Built-in tools allow comparison of:
  • Performance across models - Side-by-side capability analysis
  • Cost efficiency - Cost per quality unit analysis
  • Speed benchmarks - Response time comparisons

    • Future Updates

    The mistral.yaml file will be updated as new Mistral models are released. Updates include:

  • New model additions - Latest model releases
  • Performance updates - Refined task scores based on testing
  • Pricing updates - Current API pricing information
  • Capability additions - New features and capabilities

    • Support and Troubleshooting

    Common Issues

  • Model not found - Check API name spelling
  • Context length exceeded - Use models with larger context windows
  • Function calling not supported - Use models that support function calling
  • High costs - Switch to more cost-effective models

    • Best Practices

  • Start with small models - Test with mistral-small-latest first
  • Monitor costs - Use cost tracking features
  • Choose appropriate context length - Match model to input size
  • Use specialized models - Use codestral- for code tasks

    • This configuration ensures optimal model selection and performance within the MindX system while maintaining cost efficiency and quality standards.

    All DocumentsDocument IndexThe Book of mindXImprovement JournalAPI Reference