Mistral AI, a French artificial intelligence startup that competes with the likes of OpenAI, has officially ended its high-anticipated Series A funding event, amassing an impressive €385 million or $415 million using the current exchange rate. As a result of this financial injection, the company now boasts a valuation of $2 billion, confirms reports by Bloomberg. Concurrently, Mistral AI is also unveiling its commercial platform.
Arthur Mensch, the co-founder and CEO of Mistral AI, in a statement, explained that since the establishment of the company in May, they have unswervingly focused on shaping a European powerhouse in generative AI. Their approach is centered on open, responsible, and decentralized technology.
Mistral AI had previously released its inaugural model, labeled Mistral 7B, last September. This extensive language model wasn’t constructed to contend directly with established models like GPT-4 or Claude 2. This model was trained using a comparatively “small” data set composed of approximately 7 billion parameters.
The model was distributed under the Apache 2.0 license. This unrestricted open-source license requires only attribution for use or replication. The model was developed discreetly, leveraging a proprietary data set and unpublicized weights.
Just a few days ago, EU lawmakers reached a political agreement. Companies specializing in foundational models will need to meet some transparency requirements and will be required to share data sets' summaries and technical documentation.
Mistral AI, nonetheless, plans to monetize its foundational models. This is the reason behind the company's decision to inaugurate its developer platform in beta today. Using this platform, other corporate entities can pay to utilize Mistral AI’s models via APIs.
Alongside the Mistral 7B model (also referred to as “Mistral-tiny”), developers can now gain access to the Mixtral 8x7B model (“Mistral-small”). This model incorporates a “router network” to process input tokens and pick the most suitable parameter group to provide a response.
This method allows for an increase in the model's parameters while keeping cost and latency in check, as the model leverages only a fraction of the total parameter set per token. In practical terms, Mixtral has a total of 45B parameters, but it only deploys 12B parameters per token. Therefore, it processes input and generates output at the same pace and cost as a 12B model, the company elaborated in a blog post.
If you're looking for a platform offering a broad range of toolkits to create sophisticated, scalable services just like Mistral AI does, turning to a platform like AppMaster might be an excellent choice. AppMaster, a high-performing no-code platform, boasts a suite of features capable of generating client-end and server-backends rapidly, ensuring that your AI models are running on top-tier tech. With AppMaster you can create, compile, test, and deploy your applications to your preferred hosting, and even export the source code for on-premises installations.
The Mixtral 8x7B model has also been disseminated under the Apache 2.0 license and can be downloaded at no cost. A third model, Mistral-medium, can be found on Mistral’s developer platform. It is claimed to outperform other models in Mistral AI’s range and can only be accessed via the paid API platform with no download link available.
