1. Mistral and Microsoft’s Collaboration
Recently, Microsoft announced a collaboration with Mistral AI, which has attracted considerable attention from industry insiders. The partnership focuses on three core areas:
-
Supercomputing Infrastructure: Microsoft will support Mistral AI through Azure AI supercomputing infrastructure, providing top-tier performance and scale for the training and inference workloads of Mistral AI’s flagship model.
-
Market Expansion: Microsoft and Mistral AI will offer advanced models of Mistral AI to customers via model-as-a-service (MaaS) in Azure AI Studio and the Azure Machine Learning model catalog. In addition to OpenAI models, the model catalog provides various options for open-source and commercial models. Today, Microsoft Azure MACC can be used to purchase Mistral AI’s models.
-
AI Research and Development: Microsoft and Mistral AI will explore collaboration on training purpose-specific models for particular clients (including workloads for the European public sector).
The current flagship product of Mistral is Mistral-Large:
-
Fluent in English, French, Italian, German, and Spanish, with strong coding capabilities (the author has tested and found that Chinese support is also adequate).
-
Has a context window of 32k tokens, which provides excellent recall for retrieval enhancement.
-
Equipped with local function call capabilities, supporting JSON output.
-
Concise, practical, unbiased, with fully modular management control.
Currently, the Mistral-Large model can be seen on Azure Machine Learning (deployable in the Central region of France and the East Coast of the US):
2. Mistral AI’s Own SaaS Solution
https://mistral.ai/
The multilingual dialogue assistant Le Chat, based on the Mistral model, serves as a conversational entry point for interacting with various Mistral AI models. It provides an educational and engaging way to explore Mistral AI’s technology.
Le Chat can operate in the background using Mistral Large or Mistral Small, or a prototype model called Mistral Next. For enterprises, Mistral AI is launching Le Chat Enterprise.
Let’s first verify Le Chat’s capabilities against Azure OpenAI GPT4-0125 (AOAI).
We will randomly ask three questions.
Question 1: Please list three common states of water and briefly describe the conditions for their transitions.
Mixtral:
AOAI:
It seems that the capabilities are comparable.
Question 2: Imagine a brand new mode of transportation that can solve a major problem faced by current vehicles. Please briefly describe the characteristics of this transportation mode and how it addresses the issue.
Mistral:
AOAI:
It seems that the capabilities are comparable.
Question 3: Teacher Wang has a box of pencils. When divided among 2 students, there is 1 left over; when divided among 3 students, there are 2 left over; when divided among 4 students, there are 3 left over; and when divided among 5 students, there are 4 left over. What is the minimum number of pencils in this box?
Mistral:
AOAI:
AOAI answered correctly, while Mistral answered incorrectly.
From the tests, AOAI has stronger mathematical abilities, but Mistral-Large, as a new AI model, already demonstrates considerable capabilities, particularly in supporting Chinese Q&A effectively.
3. Classification of Mistral AI’s Underlying Models
https://mistral.ai/technology/#models
Mistral AI’s models are classified into two main categories: optimized and open-source.
Currently, there are no open-source options for the optimized models:
Optimized models can be subscribed to on a monthly payment basis and accessed via API keys.
The open-source models mainly include Mistral 7B and Mistral 8*7B
The open-source models have previously been available on Hugging Face:
Both closed-source and open-source models of Mistral are available on Azure Machine Learning Studio:
In Azure Machine Learning Studio, models can be deployed with a single click and then called:
In previous articles, I also conducted some validations based on Mistral 8*7B. Mistral-8x7B is an implementation of the SMoE architecture. Innovation.
Mistral-8x7B: SMoE’s 8-cylinder engine – PyTorch Learning Series 45
Running Mixtral-8x7B with mistral-offloading – PyTorch Learning Series 46
In the future, I will continue to validate solutions related to Mistral, and I believe that Microsoft’s platform will offer more collaboration opportunities with Mistral.