Mistral AI has released its most advanced language model to date, Mistral Large. Mistreal boasts that the text generation model offers top-tier reasoning capabilities, making it suitable for complex multilingual reasoning tasks, including text understanding, transformation, and code generation.
Mistral Large introduces several new capabilities and strengths that set it apart from its predecessors:
1. Native Multilingual Fluency: The model is fluent in English, French, Spanish, German, and Italian, demonstrating a nuanced understanding of grammar and cultural context in each language.
2. Extended Context Window: With a 32K tokens context window, Mistral Large enables precise information recall from large documents.
3. Precise Instruction-Following: Developers can design their own moderation policies using Mistral Large's precise instruction-following capabilities, as demonstrated by the system-level moderation of le Chat.
4. Function Calling: Mistral Large is natively capable of function calling, which, combined with the constrained output mode implemented on la Plateforme, enables application development and tech stack modernisation at scale.
Mistral Large achieves strong results on commonly used benchmarks, positioning it as the world's second-ranked model generally available through an API (next to GPT-4). The model demonstrates powerful 'reasoning' capabilities, outperforming other top-leading LLM models on standard benchmarks such as MMLU, HellaSwag, Wino Grande, Arc Challenge, TriviaQA, and TruthfulQA.
In terms of multilingual capacities, Mistral Large strongly outperforms LLaMA 2 70B on HellaSwag, Arc Challenge, and MMLU benchmarks in French, German, Spanish, and Italian.
The model also shows top performance in coding and maths tasks, as evidenced by its results on popular benchmarks like HumanEval, MBPP, Math, and GSM8K.
Alongside Mistral Large, Mistral AI has released Mistral Small, an optimised model designed for latency and cost.
Mistral Large and Mistral Small support JSON format mode, which forces the language model output to be valid JSON. This functionality enables developers to interact with the models more naturally and extract information in a structured format.