Soon after Google’s Gemini Ultra announcement, France-based AI startup Mistral unveiled its largest language model to date, dubbed Mistral Large. The model stands out for its aim to embrace the rich tapestry of European languages and cultural nuances. In line with industry expectations, the model remains proprietary and not available as open-source software.
Dubbed ‘Mistral Large,’ this cutting-edge model boasts a 32,000 context window, setting it up as Europe’s formidable challenger to OpenAI’s renowned GPT-4. This strategic move underscores Mistral’s ambition to carve out a significant niche in the global AI landscape.
Mistral articulates that this innovative model excels in a variety of complex multilingual operations. These include but are not limited to, understanding texts, converting text across formats, and generating computer code.
Mistral Large has demonstrated impressive performance across widely recognized industry benchmarks, securing its position as the second top-performing model accessible through API services—just behind GPT-4—in the prestigious MMLU language comprehension benchmark.
While Google’s Ultra model surpasses its capabilities, it remains out of reach for general API access. Additionally, Mistral Large offers a cost advantage, being approximately 20% more affordable than the GPT-4 Turbo option. Here are the results.
As mentioned earlier, Mistral’s latest language model is meticulously engineered to grasp the nuanced cultural distinctions and linguistic variety prevalent across European nations. It boasts fluency in multiple languages, including English, French, Spanish, German, and Italian. Mistral highlights the model’s adeptness at comprehending both the grammatical intricacies and the cultural contexts inherent to these languages.
Developers should find it easy to shape their moderation policies with Mistral Large since the model has the ability to follow precise instructions. It can also be used with other tools and apps for more complex international since it has support for “function calling.”
Mistral Small
In addition to Mistral Large, the French company has also introduced Mistral Small, which is roughly the size of Mixtral 8x7B and is meant to focus on low latency and cost-effectiveness. Unlike Mistral Large, Mistral Small is fully open-source.
Here is how all Mistral models compare with each other through benchmarks.
Mistral has made the Large model accessible through its proprietary platform “La Plateforme” as well as Microsoft’s Azure AI, offering API connectivity. The company reports that beta customers have already achieved successful outcomes using the model.
Additionally, Mistral is providing a chance to explore the capabilities of the model at no cost in a chat interface through its newly unveiled assistant. This trial also extends to the recently introduced Small model, allowing for comprehensive testing.
