The Mistral AI chatbot offers a range of complete features for a large number of use cases. A true sovereign alternative to ChatGPT.
Mistral maintains its lead in the European race for generative AI. After launching Le Chat last February, its assistant based on its own artificial intelligence models, the French start-up is offering new tools this November to compete with ChatGPT. Web search, Canvas mode, advanced vision, image generation… But do the new features announced on November 18 raise Le Chat to the level of the American AI giant?
A simple and functional interface
For its Chat update, Mistral has changed its interface very little. The latter always remains very sober and functional. Mistral has simply streamlined its interface a little further. The left column, with the conversation history, now appears when hovered with the mouse. The conversational space now occupies all available space. Major new feature: it is no longer possible to select the Mistral model to use in the conversation. The system appears to use multiple Mistral AI models depending on the complexity of the query. A development which likely allows Mistral to reduce its overall inference costs.
For multimodality, however, we know that Mistral uses its new flagship: Pixtral Large. The latter is, according to benchmarks, excellent in the analysis of documents (DocVQA: 93.3), complex diagrams and graphs (AI2D: 93.8) and even outperforms GPT-4o and Claude 3.5 Sonnet. On the other hand, the model falls slightly short on complex multimodal reasoning (MMMU: 64). Clearly, the model is perfect for analyzing documents with many graphics, a common use case in business.
We were able to test the model on several queries involving complex graphics and visual elements. The model offers an analysis and interpretation of elements as qualitative as GPT-4o or Claude 3.5 Sonnet.
Example below with analysis of a Tesla stock chart since the start of the year. We ask the AI to generate its analysis and interpret the news with web search. Mistral gives a good analysis but does not correlate the rise in action in November with the election of Donald Trump. On the other hand, ChatGPT makes the connection between the two. OpenAI's chatbot also offers more compact analysis with less robotic text. Mistral still sprinkles his text with numerous logical connectors and a conclusion typical of an LLM.
A very (too) basic web search
For the web search part, Mistral AI seems to be based on Bing. It is very simple to use but the information presented by Le Chat is only based on a few sources, often two. For its part, ChatGPT bases – very often – its answers on 5 to 6 different sources. Likewise, the Chat web search is very basic and often provides dated information. The interest of the tool is limited to providing the model with knowledge on very precise queries. For its part, ChatGPT brings truly new elements. This radical difference in quality is largely explained by the search algorithms used, OpenAI recently having its own web indexing engine.
Example below with the query “What are the latest news in the LLM sector?” Chat only gives fairly old results.
Canvas mode, assistant configurator
Like ChatGPT, Le Chat now has a Canvas mode. The latter is very practical and allows you to rework text or code in blocks. The model no longer needs to regenerate all the text or all the code. Faster and more fluid for the user and more economical in inference for editors. The operation is essentially the same between Mistral AI and ChatGPT. The only difference is in the interface. ChatGPT offers a simpler to use window which allows you to select the code or text to modify when Chat only offers a window to view the modifications. With Chat, it is therefore necessary to explicitly indicate in the prompt the block or line to change. Finally, ChatGPT's versioning system is much simpler to use (in historical form).
On the assistant side, ChatGPT and Le Chat both offer the possibility of creating your own bots. The Mistral configurator is a little limited: it only allows you to configure the model used, the prompt, and also – unlike OpenAI – the temperature of the model. The configurator also allows you to add. Examples of questions/answers for the model to draw inspiration from.
At OpenAI, ChatGPT allows you to configure the prompt, add a documentary base, add tools (web search, code interpreter, etc.) and even more useful, configure external action APIs.
Also note that configuring agents at Mistral requires exiting Chat and using the developer platform. A little heavy to use.
Image generation: an unbalanced match
For image generation, ChatGPT is based on its aging Dall-E 2 model while Chat is based on FLUX1.1 [pro]a model developed by the German Black Forest Labs. The integration of the latter in Chat seems quite limited. It is not possible to generate images in any format other than 1024 × 768 (0.79 MP). For its part, Dall-E in ChatGPT can generate a wide variety of formats, up to 1792×1024 (1.84 MP). A more decent maximum size for creative use. In terms of visual quality, FLUX1.1 [pro] wins the match. The Black Forest model is excellent in the majority of styles and remains state of the art in photorealism.
Compliance: Le Chat advantage
Le Chat also stands out in terms of compliance and data sovereignty. Unlike ChatGPT which hosts its servers in the United States, Le Chat processes and stores all data in Europe. This configuration allows French and European companies to fully comply with their regulatory obligations in terms of data protection, while avoiding the risks linked to the American Cloud Act. A strong argument for organizations handling sensitive data, particularly in the public, banking or health sectors.
Chat: a credible alternative to ChatGPT
Mistral AI Chat stands out as a credible alternative to ChatGPT for French companies. With a clean interface and functionalities almost similar to those of OpenAI – document analysis, code generation, Canvas mode, creation of assistants – the sovereign solution meets the daily needs of many organizations. A professional subscription, scheduled to be launched soon, should further enrich the offer.
For the most demanding users, certain limitations nevertheless persist. Web search remains basic, Canvas mode lacks ergonomics and the assistant configurator does not yet offer all the advanced features of ChatGPT. OpenAI also offers additional features that are very useful on a daily basis, such as memory or connection to third-party storage services. The absence of an advanced reasoning model in Mistral – essential for developing agentic capabilities – and the choice not to develop its own image generation model may also be surprising. However, this strategy focused on the excellence of LLMs is bearing fruit, and new features should enrich the platform in
