Magistral + Unlimited Free APIs + Local Usage : This FULLY FREE Coder Reasoning Model is PRETTY GOOD

Magistral: Mistral's New Reasoning Model - Detailed Summary

Key Concepts:

• Magistral Small & Medium: Mistral's new reasoning models.
• Reasoning Model: AI model focused on multi-step logic, structured calculations, and decision trees.
• Transparency: Model's ability to show its thought process.
• Multilingual Capability: Ability to reason natively in multiple languages.
• Lasha API: Platform for accessing Magistral Medium.
• Open-source: Freely available model with accessible code.
• Reinforcement Learning: Training method used to improve reasoning abilities.
• Multimodal: Ability to process different types of data (e.g., text and images).
• Token Throughput: Speed at which the model processes text.
• Mistral Code: AI coding assistant by Mistral in collaboration with Continue.
• Dart: AI-native project management tool.
• Rode, Klein, Kilo: Platforms for using AI models.
• Ollama: Tool for running local models.

1. Introduction to Magistral

• Mistral AI has launched Magistral, their first reasoning model, available in two versions:
  • Magistral Small: Open-source model with 24 billion parameters.
  • Magistral Medium: Aimed at enterprise users.
• The primary focus is on reasoning and transparency, enabling the model to handle multi-step logic, structured calculations, and decision trees.
• Unlike general-purpose models, Magistral aims to make its thought process clear, showing each step in arriving at an answer.

2. Benchmarks and Performance

• Magistral Medium achieved a score of 73.6% on the Amy 2024 benchmark. With Majority Voting, it reached 90%.
• Magistral Small scored 70.7% and 83.3% respectively.
• These scores are considered decent, especially for an open-source model.

3. Multilingual Capabilities and Speed

• Magistral can reason natively in multiple languages, including English, French, Spanish, German, Italian, Arabic, Russian, and simplified Chinese.
• This allows users to receive step-by-step answers in their own language.
• Optimized for speed, especially when used in Lasha with the flash answers feature, claiming up to 10 times faster token throughput compared to competitors.

4. Applications and Use Cases

• Magistral is versatile and suitable for:
  • Legal research
  • Software development
  • Creative writing
• Ideal for scenarios requiring accurate multi-step reasoning and transparent answer derivation.
• Provides traceable reasoning for regulated industries like legal, finance, healthcare, and government, aiding in compliance.

5. Technical Details and Availability

• Magistral Small is open-weight and available under the Apache 2.0 license for download and self-hosting.
• Magistral Medium is accessible through the Lasha API and platforms like Amazon Sage Maker, with more cloud options planned.
• The research paper detailing the methodology behind Magistral is also available.
• Both models are based on Mistral Small 3 and Mistral Medium 3, enhanced with reinforcement learning for improved reasoning.
• Similar to Deepseek, Magistral models are multimodal, making them unique among open reasoning models.

6. Accessing and Using Magistral

• Magistral Small can be used via Olama due to its open-source nature.
• Magistral Medium is available for free on the Lasha interface and through their API platform's free tier (with data sharing), offering:
  • Approximately one request per second.
  • A generous limit of 500,000 tokens per minute.
  • Up to 1 billion tokens per month.
• The Lasha platform collaborates with Cerebras for fast inference.
• Magistral Small is also available on Open Router.

7. Mistral Code and Integration with Development Tools

• Mistral Code, a new AI coding assistant developed with Continue, is currently available only for enterprise users.
• Integration with platforms like Rode, Klein, and Kilo is demonstrated:
  • Users can set up profiles on these platforms, select Mistral or Open Router, and enter the necessary details.
  • These platforms support tool calling.
  • Kilo offers $20 free credit for using paid models.
• Example: Generating a 3.js car driving game using Rode. The model successfully creates the pages and performs tool calling.

8. Dart: AI-Native Project Management Tool (Sponsored)

• Dart is an AI-native project management tool for managing tasks, creating boards, and organizing projects.
• AI features include:
  • Generating tasks through brainstorming and project planning.
  • Duplicate detection.
  • Assigning tasks to AI agents for completion.
  • Integration with AI clients or coders via its MCP server.
  • Integration with Claude, ChatGPT, and more.
• Most features are free, with a $8 subscription available for additional features.

9. Performance Comparison and Future Expectations

• Magistral's performance is comparable to the previous generation Deepseek.
• Magistral Small is considered one of the best local models currently available.
• There is an expectation for Mistral to develop a larger Magistral model to achieve state-of-the-art performance.

10. Conclusion

• Mistral's Magistral models represent a significant step in reasoning AI, offering transparency, multilingual capabilities, and speed.
• The open-source availability of Magistral Small and the free tier access to Magistral Medium on Lasha provide opportunities for experimentation and practical use.
• While there is room for improvement, particularly in achieving state-of-the-art performance, Magistral is a promising development in the field of AI reasoning.