Mistral AI releases Pixtral Large and upgrades Le Chat

Two models in one day

On 18 November 2024 Mistral AI announced Pixtral Large, a 124-billion-parameter open-weight multimodal model built on Mistral Large 2^[1]. Multimodal means the model takes images as well as text as input, not text alone. Open-weight means the trained parameters are published so that others can download and run the model; that is a different thing from publishing the training data or the training code. Parameter count is a rough indicator of model size, not of how well a model does on a given task.

The same day brought Mistral Large 24.11, an update to the text model with better handling of long contexts and improved function calling^[1]. Long-context handling is about keeping track of information across a long input — a whole document, or a long conversation — instead of only the most recent paragraphs. Function calling is the ability to return a structured request to an outside tool, such as a search engine or a database, rather than only producing prose. Mistral's changelog records the two releases under the identifiers mistral-large-2411 and pixtral-large-2411^[3].

The MathVista score

Mistral reported a score of 69.4% on MathVista for Pixtral Large^[1]. MathVista is a benchmark built around mathematical questions whose information is partly carried by an image — a chart, a diagram, a figure — so a model has to read the picture and then reason over what it shows, rather than match patterns in text alone. A benchmark score is a measurement on one particular test set, and says little by itself about how a model behaves on other kinds of material.

Le Chat gets new features

Le Chat, the company's chat assistant, received several additions at the same time: web search, Canvas, image generation and agents^[2]. Image generation runs on Flux Pro, a model from Black Forest Labs^[2]. Canvas is a separate editing area beside the conversation, where a user works on a piece of text or code together with the model instead of only exchanging messages. Agents let the assistant carry out a multi-step task rather than answer a single question. In Le Chat, Pixtral Large handles documents and images^[2] — the part of the product where someone uploads a file or a picture and asks about it.

Mistral also said that faster Le Chat responses were powered by speculative editing^[3]. Speculative techniques use a cheaper, faster draft to propose several tokens ahead, and the main model then checks that proposal; wherever the draft guessed right, generation moves ahead faster than producing every token with the large model alone, and wherever it guessed wrong, the result is simply rechecked.

What the day put together

The announcement tied a model release and a product release to the same date. On one side is Pixtral Large, an open-weight multimodal model whose weights carry a dated version identifier^[1][3]; on the other is Le Chat, where the same model is the one named for document and image handling, and where web search, Canvas, image generation and agents sit next to it^[2]. For someone using the chat product, the visible change is that documents, pictures and generated images belong to the same conversation; for someone who wants to run the model, what appeared that day was a set of downloadable weights with a specific version number to point at.