Mistral launches Large 2 with 123 billion parameters

On July 24, 2024, Mistral AI introduced Mistral Large 2, a 123-billion-parameter model with long context, multilingual support and weights available for research. Mistral’s technical announcement publishes capabilities, licence, evaluations and access; each answers a different question and should not be compressed into “more powerful”.

The launch matters for two reasons. Mistral is not just updating the flagship model it introduced in February; it is also releasing its weights under a research license and significantly lowering API access prices. That combination is unusual at the high end of the market, where the most capable models typically operate exclusively as closed services.

A large model for code and lengthy documents

Mistral Large 2 supports a context window of up to 128,000 tokens. In simple terms, it can receive and connect very long texts within a single conversation: technical documentation, code repositories, contracts or collections of reports. Supporting that volume does not guarantee that a model will reason well about every detail, but it removes a common practical limitation when working with large files. Source

The company is focusing on code generation and says the model supports more than 80 programming languages. It also strengthens function-calling capabilities, the mechanism that enables an assistant to decide whether to use an external tool — for example, query a database or run an operation in a program — rather than simply generate text. Source

In its evaluations, Mistral gives Large 2 a 92% score on HumanEval, a test of small Python programming problems; 84% on MATH, which focuses on mathematical exercises; and 84% on MMLU, a broad test of knowledge and comprehension. These are useful benchmarks for comparing models, although they do not replace testing with each company’s real-world data, code and workflows. Source

Multilingualism as a European priority

The language improvements are among the most significant elements of the announcement. Mistral Large 2 is designed to work in English, French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese and Korean, among other languages.

Most large models are trained on a far greater share of English-language content. That can lead to less reliable answers, poorer summaries or misunderstood instructions when they are used in other languages. For European companies that write, support customers or analyze documents in multiple languages, multilingual quality is not a secondary feature: it determines whether a model can be integrated into the workflow without first translating the information.

Mistral had already made French and other European languages one of its defining features. Large 2 expands that strategy at a time when the market is concentrating around US and Chinese labs, with models that are increasingly expensive to train and operate.

Weights available, but not open-source software

Mistral Large 2’s weights can be downloaded for research under the Mistral Research License. This allows researchers to study and test the model on their own infrastructure, which is not possible with fully closed services.

It is important to distinguish this from an open license without commercial restrictions. The research license does not automatically grant permission to use the model in commercial products or services. Companies seeking that use must contact Mistral to obtain the applicable terms or use the service hosted by the company and its partners.

The model is available through La Plateforme, Mistral’s API, at a price of $2 per million input tokens and $6 per million output tokens. It is also coming to platforms including Amazon Bedrock, Azure AI Studio, Google Cloud Vertex AI, IBM watsonx and NVIDIA NIM. Source

The price cut is significant: the previous Mistral Large launched at rates of $8 per million input tokens and $24 per million output tokens. For applications that process large volumes of text or continuously generate code, the per-query cost can determine which model is viable. Source

The decisive test will be whether those improvements hold up beyond benchmarks. Mistral Large 2 is entering a segment where quality matters, but so do privacy, ease of deployment and a predictable bill. The ability to research it locally and use it across multiple clouds gives customers and developers more options for comparing those factors.

A model record prevents false comparisons

A commercial name is not enough to reproduce a result. The minimum record contains version identifier, date, licence, weight format, quantisation, chat template, context used, generation parameters and hardware. Two teams can say they run Large 2 and obtain different latency, memory use and answers because those conditions differ.

Advertised context is input capacity, not perfect memory. A test distributes facts across the beginning, middle and end of a document, adds contradictory information and asks for the exact source. Measure retrieval, fidelity and the tendency to fill gaps. Supplying a whole file saves segmentation only if the model consistently finds what matters.

Benchmark, language and task must share an axis

An average coding score does not demonstrate performance in a company repository. Preserve the language, dependencies, tests and acceptance criterion, then execute generated code. For multilingual assessment, write cases originally in each language rather than translating an English set, and inspect instructions, nuance and local terminology.

Downloadable weights increase inspectability, but a research licence is not commercial permission and reveals no training data. Separate technical access, usage rights, documentation and modification. “Open” without a qualifier hides those differences.

The transferable skill is to compare models through a stable record and your own task. Calculate quality, latency, memory, licence and cost per accepted result. If one variable changes, the same comparison is no longer being repeated even when the model name remains.

Useful cost ends with an accepted answer

Per-token pricing helps budget traffic but does not show how much work the system completes. A comparison preserves the same requests, counts retries, rejected answers, tools and human review, then divides total spending by outputs that pass the criterion. A cheap model requiring repetition may cost more per finished task.

Local deployment adds another ledger: memory, accelerators, energy, operations staff and idle time. An API turns some of those costs into a tariff while transferring control to the provider. Neither option wins by definition; the choice depends on volume, data sensitivity, tolerance for change and the team’s ability to operate the system.

A small, repeatable protocol

Select real cases, pin a version and run them repeatedly with recorded settings. Inspect accuracy, citations, format and dangerous failures; measure latency and cost alongside quality. Repeat after changing the model, prompt template or provider. The announcement becomes a testable hypothesis and the comparison survives the next commercial name.

Before deciding, publish the record and ask another team to reproduce a sample. Disagreements reveal hidden dependencies: permissions, library versions, service limits or review decisions. If the second team cannot repeat the result, the figure still describes a demonstration rather than an operating capability. Recording rejected cases also prevents the report from being optimised by hiding the most expensive requests. The protocol becomes more useful when another team can challenge it from the same evidence.