Mistral said on 24 July 2024 that it had launched Mistral Large 2, a 123-billion-parameter model designed for single-node inference with long-context applications.
The model has a 128k context window, supports dozens of languages and more than 80 coding languages, and achieved 84.0% accuracy on MMLU in its pretrained version. Mistral said it performs on par with leading models such as GPT-4o, Claude 3 Opus and Llama 3 405B on code and reasoning work.
Training emphasised reducing hallucinations, better instruction following and multi-turn conversation, plus enhanced function calling and retrieval for parallel and sequential tool use. Weights for the instruct model are on Hugging Face under the Mistral Research License for non-commercial use; commercial self-deployment needs a Mistral Commercial License.
On la Plateforme the model is available as mistral-large-2407 and can be tried in Le Chat. Fine-tuning capabilities were extended to Mistral Large, Mistral NeMo and Codestral.
Cloud availability at launch expanded to Google Cloud Vertex AI via a managed API, alongside Azure AI Studio, Amazon Bedrock and IBM watsonx.ai.