ContextQMD
Back to Promptfoo

mistral (Mistral AI Chat Models)

examples/mistral/README.md

0.121.184.6 KB
Original Source

mistral (Mistral AI Chat Models)

This example demonstrates Mistral AI's chat models, including Magistral reasoning models, current multimodal models, and shows how to use Mistral models for evaluation grading and embeddings.

You can run this example with:

bash
npx promptfoo@latest init --example mistral
cd mistral

Environment Variables

This example requires:

What This Example Shows

  • Mathematical Reasoning: AIME2024 competition problems with Magistral Medium
  • Model Comparison: Compare Mistral's different model capabilities
  • Reasoning Models: Showcase Magistral Medium (native reasoning) vs. Mistral Small 4
  • Chat Capabilities: General conversation and task completion
  • Mistral-powered Evaluation: Use Mistral models for grading instead of OpenAI
  • Mistral Embeddings: Use Mistral's embedding model for similarity checks

Models Demonstrated

Reasoning Models

  • Magistral Medium (magistral-medium-latest → magistral-medium-2509): Native reasoning model ($2/$5 per 1M tokens, 128k context) — the reasoning showcase in these examples.

Mistral folded Magistral Small into Mistral Small 4: the magistral-small-latest alias now resolves to mistral-small-2603 (a hybrid model, $0.15/$0.60 per 1M), so these examples use the canonical mistral-small-latest id. Enable Small 4's reasoning mode with reasoning_effort: high. The standalone magistral-small-2509 snapshot is deprecated (retires 2026-07-31).

Chat Models

  • Mistral Medium 3.5 (mistral-medium-latest → mistral-medium-2604): Frontier agentic/coding multimodal model ($1.50/$7.50 per 1M, 256k context)
  • Mistral Large 3 (mistral-large-latest → mistral-large-2512): General-purpose multimodal model ($0.50/$1.50 per 1M, 256k context)
  • Mistral Small 4 (mistral-small-latest → mistral-small-2603): Hybrid instruct/reasoning/coding model ($0.15/$0.60 per 1M, 256k context)

Evaluation Models

  • Grading: Uses mistral-large-latest for LLM-as-a-judge evaluation
  • Embeddings: Uses mistral-embed for semantic similarity checks

Key Features Demonstrated

  • Multi-model comparison: Compare performance across different Mistral models
  • Reasoning capabilities: Step-by-step problem solving with Magistral models
  • Cost optimization: Balance performance vs. cost across model tiers
  • Self-evaluation: Use Mistral models to grade their own outputs
  • Semantic similarity: Mistral embeddings for content comparison

Running the Example

bash
# Set your API key
export MISTRAL_API_KEY=your_api_key_here

# Run the evaluation
promptfoo eval

# View results in the web UI
promptfoo view

Configuration Highlights

This example showcases several advanced promptfoo features:

  • Provider overrides for grading and embeddings
  • Multiple assertion types including llm-rubric and similarity
  • Cost tracking across different model tiers
  • Mixed scenarios from simple chat to complex reasoning

The evaluation uses Mistral models end-to-end, providing a comprehensive view of their ecosystem capabilities.

Available Configurations

This example includes multiple configuration files for different use cases:

Mathematical Reasoning

  • promptfooconfig.aime2024.yaml - Advanced mathematical competition problems (AIME2024 dataset)
  • promptfooconfig.reasoning.yaml - Step-by-step logical problem solving

Model Capabilities

  • promptfooconfig.comparison.yaml - Compare reasoning across all Mistral models
  • promptfooconfig.code-generation.yaml - Multi-language programming with Codestral
  • promptfooconfig.multimodal.yaml - Vision and text processing with current Mistral multimodal models

Advanced Features

  • promptfooconfig.tool-use.yaml - Function calling and tool integration
  • promptfooconfig.tool-routing.yaml - End-to-end QA for tool-only, mixed content+tool_calls, file-based tools, and plain chat output
  • promptfooconfig.json-mode.yaml - Structured JSON output generation
  • promptfooconfig.yaml - Main example with evaluation using Mistral models

Run any specific configuration:

bash
npx promptfoo@latest eval -c promptfooconfig.aime2024.yaml  # Mathematical reasoning
npx promptfoo@latest eval -c promptfooconfig.comparison.yaml  # Model comparison

Additional Resources