The artificial intelligence technology industry has recently added another promising participant – the French startup Mistral AI. In our article, we will tell you how this project arose and is developing, what products it has released, and why it is considered a serious competitor to OpenAI.
What is Mistral AI
Mistral AI is a generative AI software company founded in 2023 in France. It was founded by three former employees of Google DeepMind and Meta: Arthur Mensch (CEO), Guillaume Lample (Chief Scientist), and Timothée Lacroix (CTO). In June 2023, the firm received its first funding of €105 million from a group of investors including Lightspeed Venture Partners, JCDecaux, Eric Schmidt, and Xavier Niel. After that, its estimated value was 240 million euros.
In September 2023, the development team presented its first product, the Mistral 7B language model, distributed under the free Apache 2.0 license. In December 2023, Mistral AI raised another 385 million euros in its second investment round. The venture fund Andreessen Horowitz, BNP Paribas, and Salesforce participated in it. Shortly after this, a more powerful version of the open-source neural network Mixtral 8x7B was released, which outperformed Meta's Llama 2 70B in a number of tests.
In February 2024, Microsoft announced a partnership with a promising AI startup. In addition to its $16.3 million financial investment, it added Mistral language models to its Azure cloud service. Amazon has also integrated the Mistral 7B and Mixtral 8x7B neural networks into the Amazon Bedrock platform. In April, it became known about negotiations between Mistral AI founders and investors to raise an additional 500 million euros. This should increase its capitalization to 5 billion euros.
AI Products of the Company
Mistral AI produces open source (under the free Apache 2.0 license) and commercial language models. We invite you to briefly familiarize yourself with the representatives of both groups.
Open LLMs:
- Mistral 7B. The company's first neural network was trained on a dataset of 7 billion parameters, but its capabilities correspond to models with 30 billion parameters. Works with English text and program code. It is easy to configure and quick to deploy. Processes large volumes of data with minimal resource requirements.
- Mixtral 8x7B. This is a more powerful open-source model. It appeared in December 2023. To output data, it uses 12 billion parameters out of a potentially available 45 billion. This version of Mixtral surpassed GPT 3.5 (175 billion parameters) and Llama 2 (70 billion parameters) in most tests. Works with code and text in English, German, French, Spanish, and Italian.
- Mixtral 8x22B. The largest and most powerful open-source model to date from Mistral AI. Its release took place in April 2024. It uses 39 billion parameters out of a potential 141 billion available. Ideal for generating large amounts of text and analyzing multipage documents in multiple languages.
Commercial LLMs:
- Mistral Small is the youngest neural network in the line. It copes well with simple text generation and classification tasks. This LLM is less resource-intensive for low-latency workloads. It supports program code and multilingual text.
- Mistral Large – the largest and most productive Mistral AI model. It is optimally suited for solving problems of generating text and code of high complexity. According to the results of individual tests, this neural network took second place after GPT-4.
- Mistral Embed is a special LLM for transforming text into computer-recognizable numeric representations (“embeddings”). It does a good job of text classification and sentiment analysis. Currently, it only supports English.
The company's product list is complemented by the free AI chatbot Le Chat, available at http://chat.mistral.ai/chat. It is similar in functionality to ChatGPT, Claude, and Gemini. It communicates with users in the same way, answers their questions, writes, and edits text. When starting the bot, you have the opportunity to select the model loaded into it. Mistral Large (for productivity), Mistral Small (for speed), or the Mistral Next prototype model, tailored for short and succinct answers. Today, the company is working on creating a commercial version of the chatbot for corporate clients.
Moreover, Mistral AI released La Plateforme, a platform for third-party developers and teams. It allows you to create new AI products based on free and commercial LLM Mistral.
How to Use Mistral AI and How Much It Costs
You can access Mistral’s open and commercial neural networks on the company’s official website. The functionality for launching them is located on La Plateforme (https://console.mistral.ai/). There are also tools for developing and deploying AI applications. The platform allows developers to train and tune models to their specifications and then integrate them into any third-party software. The data on it is encrypted both during storage (AES256) and during transmission (TLS 1.2+). In addition, the platform offers guides and forums with content useful for AI enthusiasts.
There are other ways to use LLMs Mistral. For example, you can download and install them on your local computer or access them through the API. When choosing the second option, developers need to register on https://mistral.ai/ and generate API keys on the page https://console.mistral.ai/api-keys/. API documentation is published here: https://docs.mistral.ai/
Mistral AI API expands the functionality of third-party programs and applications by integrating the following AI tools into them:
- Generate text in streaming mode with the ability to demonstrate partial results of the model.
- Encoding text as a sequence of numbers using the "embeddings" function.
- Connecting LLM to external tools via function calls.
- Implement developer rules in the Mistral model at the system level using the "guardrailing" function.
- Selecting json_object response format using JSON mode.
Another way to deploy Mistral neural networks is through cloud services. Anyone can take advantage of their functions through cloud providers GCP, AWS, Azure, Snowflake, NVIDIA. The top-of-the-line Mistral Large model is only offered by Azure AI Studio, AWS Bedrock, and Snowflake. In addition, third-party developers and teams can run LLM in a virtual cloud or locally (on-prem). Self-service deployment provides additional customization and system management capabilities.
Mistral AI pricing is based on the pay-as-you-go model and offers a number of tariffs:
- Mistral 7B – $0.25/1M I/O tokens.
- Mixtral 8x7B – $0.7/1M I/O tokens.
- Mixtral 8x22B – $2/1M input tokens and $6/1M output tokens.
- Mistral Small – $1/1M input tokens and $3/1M output tokens.
- Mistral Large – $4/1M input tokens and $12/1M output tokens.
- Mistral Embed – $0.1/1M I/O tokens.
Capabilities of Mistral AI Language Models
Mistral LLMs are basic models. They can be trained, retrained, and deployed for a wide range of natural language processing (NLP) and code tasks. Their main abilities are:
- Content generation. Neural networks are ideal for creating texts in English and other languages (depending on the specific model). They can write a letter for an email newsletter, advertising text, a post for social networks, a resume, a short story and much more.
- Development of smart chatbots. Mistral models understand text queries and generate responses based on the data loaded into them.
- Summarizing the text. AI models analyze articles and documents, highlighting their key points in a short overview-summary.
- Text classification. Neural networks are capable of organizing texts according to various criteria. For example, sorting emails into folders based on their content.
- Code generation and addition. LLMs create code snippets based on text queries and optimize the finished code. Additionally, they check it for errors and provide recommendations for correcting them.
Mistral AI vs. GPT 4
Comparing the characteristics of Mistral AI and GPT-4 from OpenAI is of interest to many AI enthusiasts. The management of the French startup claims that Mistral Large (the senior model in their commercial product line) is comparable to GPT-4. Meanwhile, the neural network from OpenAI showed the best results in all performance tests, surpassing Mistral in solving problems in NLP, computer science, history, mathematics, and other disciplines. However, the gap between these models is relatively small. For example, according to the results of the test for language understanding under multitasking conditions, the result of GPT-4 was 86.4%, and Mistral Large – 81.2%.
Customization is an equally important criterion for comparing Mistral AI and GPT-4. According to this parameter, the French startup has a more advantageous position. It wins mainly due to its highly customizable and easily integrated Mistral AI open source models. Mistral AI products are more user-friendly and accessible to third-party developers. This is quite a strong advantage.
A comparison of LLMs would be incomplete without assessing the price factor. By this criterion, Mistral AI outperforms GPT-4. The cost of using Mistral Large via the API is $4 per million input tokens and $12 per million output tokens. The price of the new version of GPT-4 Turbo is $10 per million input tokens and $30 per million output tokens. Thus, Mistral Large costs users significantly less than GPT-4.
Conclusion
Startup Mistral AI has made a significant impact in the AI industry by releasing a variety of open and commercial LLMs within a short period of time. This company is confidently predicted to become OpenAI's main competitor, and these predictions have a high chance of coming true. Currently, Mistral AI has introduced a range of flexible and feature-rich open-source models. Furthermore, it has also brought to the market Mistral Large, one of the most powerful neural networks capable of competing with the renowned AI flagship GPT-4 Turbo.
Don't waste another minute manually transferring leads from Facebook to other systems. SaveMyLeads is a simple and effective tool that will allow you to automate this process so that you don't have to spend time on the routine. Try SaveMyLeads features, make sure that this tool will relieve your employees and after 5 minutes of settings your business will start working faster.