Mistral AI releases Mistral Large 2

What was released

On 24 July 2024, Mistral AI released Mistral Large 2, a language model with 123 billion parameters and a context window of 128K tokens^[1]. The model card describes it as a dense language model with 123 billion parameters^[2].

Three numbers carry most of the announcement. Parameters are the learned weights inside the network; the count is a rough measure of how much the model can absorb during training and how much computation each answer costs. A context window is the amount of text the model can hold in view at once — at 128K tokens, long documents, code bases or extended conversations can be handled in a single pass rather than chopped into pieces. "Dense" distinguishes the design from sparse architectures that activate only part of the network for each token: in a dense model, all parameters take part in every step.

Language and code coverage

The model supports dozens of natural languages and more than 80 programming languages^[1]. The coding claim matters because a large share of practical assistant work involves reading, writing and repairing source code across many ecosystems, not just one or two mainstream languages.

Benchmark result

On MMLU, a standard benchmark that tests knowledge and reasoning across a wide range of academic subjects, the pretrained model scored 84.0%^[1]. The model card reports the same figure for the pretrained model^[2]. The word "pretrained" matters here: the score refers to the base model before any instruction-tuning or alignment stage, so it describes raw capability rather than the behaviour of a finished chat product.

Where it was available

Weights were published on Hugging Face under the Mistral Research License, with commercial use requiring a separate licence^[1]. That arrangement lets researchers download and study the model directly while keeping commercial deployment behind an agreement. Alongside the release, the model launched on La Plateforme, Le Chat, Vertex AI, Azure AI Studio, Amazon Bedrock and IBM watsonx.ai^[1]. In practice this meant two routes to the same model: direct access to the weights for research, and managed access through Mistral AI's own services and through several major cloud platforms, where customers could call the model without hosting it themselves.

Why it matters

The release sits at the intersection of two trends. First, capability at the frontier increasingly depends on scale — parameter count, context length and the breadth of languages covered — and a 123-billion-parameter model with a 128K window was presented as a general-purpose system rather than a narrow tool^[1]. Second, the terms of access have become part of the product. Publishing weights under a research licence, while offering the same model commercially through cloud partners, is a way of distributing a model widely without giving up control over commercial use^[1].

For someone choosing a model, the practical reading is simple: the release specifies its size, its context limit, its language coverage, its benchmark score and the legal terms under which it can be used — the set of facts that determine whether a model fits a task. The MMLU figure of 84.0% gives one reference point for comparing it with other models on the same test, while the licence determines what an organisation is allowed to do with the weights afterwards^[1][2].