On Wednesday, Mistral, a Paris-based artificial intelligence (AI) company, unveiled its latest AI model, Mistral Medium 3. This model is designed for enterprise applications, boasting high performance at competitive pricing. The company claims it represents state-of-the-art (SOTA) capabilities within its category while being offered at significantly lower rates than similar models available in the market. Importantly, Mistral Medium 3 is not open-source and will not be found on public repositories like GitHub or Hugging Face.
Mistral Medium 3 Features and Pricing
In a blog announcement, Mistral provided insights into the functionalities of its new model. Earlier in March, the company launched the Small 3.1 model as an open-source option, featuring a context window capable of handling up to 128,000 tokens. Mistral Medium 3 is part of the company’s lineup aimed at enterprise users, offering paid access through cloud services and application programming interfaces (APIs).
This multimodal large language model (LLM) can be utilized in various configurations, including hybrid setups that combine cloud and edge computing, on-premises implementations, and virtual private cloud (VPC) environments. Companies are also able to further train the model with their own internal datasets, allowing for customized responses. Additionally, Mistral Medium 3 can be seamlessly integrated into existing enterprise systems and tools.
Regarding pricing, Mistral Medium 3 will be available at a cost of $0.4 (approximately Rs. 34) per million input tokens and $2 (around Rs. 171) per million output tokens. For comparison, the Claude 3.7 Sonnet model, which Mistral identifies as a competitor to Medium 3, is priced at $3 (about Rs. 257) per million input tokens and $15 (nearly Rs. 1,284) per million output tokens.
According to internal assessments conducted by Mistral, the Medium 3 model consistently outperforms competitors such as Llama 4 Maverick, GPT-4o, and Claude 3.7 Sonnet across various benchmarks, including HumanEval, ArenaHard, Math500 Instruct, and AI2D. Its performance is reported to be on par with the DeepSeek 3.1 model in several tests.
Enterprises and developers looking to utilize Mistral Medium 3 can access the model as an API through Mistral La Plateforme and Amazon Sagemaker. Additionally, the firm plans to make it available soon on platforms like IBM WatsonX, Nvidia NIM, Azure AI Foundry, and Google’s Vertex AI.
Mistral also hinted at future developments, indicating that after launching its Small and Medium models, the company is now in the process of developing an open-source Large AI model, which is expected to be released later this year.
