Large Enough

TL;DR

Mistral Large 2 features a 128k context window and supports dozens of languages and over 80 coding languages.
The model is designed for single-node inference with 123 billion parameters, optimizing throughput.
Mistral Large 2 achieves 84.0% accuracy on MMLU, setting a new benchmark for performance/cost in open models.
It was trained on a large proportion of code, outperforming previous Mistral Large and competing with models like GPT-4o and Claude 3 Opus.
Efforts were made to minimize hallucinations and improve accuracy, with the model acknowledging when it lacks sufficient information.
Instruction-following and conversational capabilities are significantly improved, with better handling of precise instructions and long multi-turn conversations.
The model excels in multilingual tasks, trained on a large proportion of multilingual data, performing strongly on multilingual MMLU.
Mistral Large 2 has enhanced function calling and retrieval skills for complex business applications.
It is available via la Plateforme, le Chat, and HuggingFace, with research and commercial licenses.
Mistral AI is consolidating its platform offerings and extending fine-tuning capabilities to Mistral Large, Mistral Nemo, and Codestral.
Partnerships with Google Cloud Platform, Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai ensure global availability.

Continue reading the original article