Mistral Large 2

What it is

Mistral Large 2 is the second-generation flagship model from Mistral AI. It was released on 24 July 2024 under version number 24.07, and the API name it is called by is mistral-large-2407 ^[1]. The model has 123 billion parameters, a size chosen so that it can be served at high throughput on a single node rather than spread across a large cluster ^[1]. It accepts a context window of 128K tokens, which means a long document can be passed in as one input instead of being cut into pieces first ^[1].

As a flagship, it was the model Mistral AI placed at the top of its range: larger than the smaller models in the same family, and intended for the hardest requests rather than for cheap, high-volume traffic. The name "Large 2" marks it as the successor to the first generation of that flagship line.

Languages and code

The model covers dozens of languages, among them French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese and Korean ^[1]. On the programming side it spans more than 80 languages ^[1]. The model card adds that the model was trained on more than 80 coding languages such as Python, Java, C, C++, JavaScript and Bash ^[4]. Broad coverage of both natural languages and programming languages matters because a general assistant is usually asked to do several kinds of work at once: translate, summarise, answer questions, and read or write code.

Benchmarks

On MMLU, a widely used multiple-choice benchmark built from academic subjects, the pretrained version of the model scored 84.0% ^[1]. The model card reports multilingual MMLU results for the pretrained model separately: 82.8% in French, 81.6% in German and 74.8% in Chinese ^[4]. Splitting a benchmark by language is a way of showing where a model is strongest and where it is weaker; the per-language figures sit below the headline number, and the gap is widest for the language furthest from the training data the model saw most of.

License and availability

The weights were published on Hugging Face under the Mistral Research License, which covers research and non-commercial use; commercial self-deployment requires a Mistral Commercial License ^[1]. That split is the practical detail most users have to check: the model can be downloaded and studied freely, but running it inside a commercial product means talking to the company first.

Beyond Mistral's own La Plateforme and Le Chat, the model is available through Google Cloud Vertex AI, Azure AI Studio, Amazon Bedrock and IBM watsonx.ai ^[1]. Offering the same model through several cloud providers lets customers keep it inside the platform they already use for data storage, billing and access control, instead of moving data to a separate vendor.

Later versions and retirement

The line did not stop at the July release. Mistral Large 24.11, published on 18 November 2024, improved long-context understanding, added a new system prompt and made function calling more accurate ^[2]. Those three changes target the parts of a model that matter most in production: handling long inputs, following instructions set at the start of a conversation, and reliably producing structured calls that let the model act through other software.

Both versions were eventually taken out of service. mistral-large-2407 was retired on 30 March 2025, and Mistral Large 2.1 (2411) was retired on 31 May 2026 ^[3].

Where it sits

Mistral Large 2 is best understood as a snapshot of what a top-of-range model looked like in mid-2024: a dense parameter count chosen for single-node serving, a long context window, and coverage wide enough to handle many languages and many programming languages in one model. Its release pattern — weights on Hugging Face under a research licence, the same model on several major clouds, and a named API version that is later updated and then retired — is also a compact picture of how model releases are packaged and maintained.