France-based AI startup Mistral has rolled out Mistral Large 2, an open-source AI model featuring 123 billion parameters. While it can be utilized for non-commercial research purposes without any licensing restrictions, a separate agreement is necessary for commercial applications.
Licensing and Availability
Mistral Large 2's open weights facilitate research and model fine-tuning by third parties. Commercial endeavors require a distinct license, and the model is accessible on Mistral's platform as well as through cloud services like Google Vertex AI, Amazon Bedrock, Azure AI Studio, and IBM WatsonX.
A day before Mistral's announcement, Meta AI unveiled its open source Llama 3.1 model. Although Llama 3.1 boasts 405 billion parameters, Mistral Large 2 holds its own with competitive performance. It supports a 128,000 token context window and is proficient in several languages, including Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean. The model excels in generating synthetic text, code, and Retrieval-Augmented Generation (RAG).
Enhanced Coding and Instruction-Following
Optimized for coding, Mistral Large 2 supports over 80 programming languages with high accuracy. It outperforms Claude 3.5 Sonnet and Claude 3 Opus in coding tasks, closely trailing GPT-4o. The model aims to minimize errors by providing careful and precise responses, enhancing its suitability for enterprise environments.
How Mistral Large 2 compares in accuracy performance compared to other major AI models
In terms of benchmarks, Mistral Large 2 scores 84 percent on the Massive Multitask Language Understanding (MMLU) measure. By comparison, Meta's Llama 3.1, GPT-4o, and Claude 3.5 Sonnet have scores of 88.6, 88.7, and 88.3 percent, respectively. Notably, domain experts achieve around 89.8 percent. The model requires about 246GB of memory at full 16-bit precision, necessitating deployment on servers with multiple GPUs.
One of Mistral Large 2's key advantages is its efficiency, requiring less memory and resulting in faster response times. This efficiency makes it a promising option for commercial applications, where it can provide quick and concise answers, potentially reducing operational costs.
Mistral is focusing on securing additional funding, developing specialized models, and bolstering partnerships with industry-heavyweight companies. The new model is acknowledged for its cost-effectiveness and strong performance, expanding the possibilities in AI applications.
