About Mistral

Mistral is a French AI start-up founded in 2023 by former Google and Meta engineers. It has quickly grown to be a notable player in the industry, especially since the release of its Mistral Large in February 2024. Given the relatively small investments it has received to date, Mistral has achieved impressive feats and proven to be a potential contender.

The French government has always touted Mistral as a proud home-grown AI start-up and a European AI force independent of US tech giants. Mistral was believed to be closely aligned with European regulations and, therefore, more suitable for tight EU AI regulations for clients such as European governments.

In February 2024, Mistral announced an agreement with Microsoft to become available on Azure and will use Microsoft’s cloud resources for further development. This pivot to its perceived European roots, which happened at a sensitive time when the EU was investigating OpenAI and Microsoft’s partnership, caused debates among European regulators. The deal is to be invested by the European Commission.

Jan 2025

Mistral released the ‘all-new’ le Chat to compete with Claude.ai and ChatGPT, and it has some neat features that make it easier for businesses to adopt. It is fast. Its Flash Answers can generate up to 1k words/sec, beating Claude.ai and ChatGPT. Apart from offering features parity with other established players, Mistral takes a slightly different approach towards openness. Unlike other vendors that use AI Assistants as a window to sell their models and services, Mistral is more open to mixing and matching models from different vendors to achieve the best results for users, which many custom AI assistants naturally do. For example, it uses Black Forest Labs Flux Ultra for image generation. It allows Enterprise users to deploy Le Chat to environments already part of the firms’ infrastructure, including on-prem and VPC, and use custom models that fit the firm’s unique requirements.

May 2024

Apart from Azure, Mistral models are available through its La Plateforme and Le Chat, which are equivalent to OpenAI APIs and ChatGPT. Mistral’s main offerings are Mistral Small, Mistral Large, and Mistral Embed. The company also released two open models, Mistral 7B and Mistral 8x7B, under the Apache 2.0 Licence.

Mistral Models & Services

Mistral Large (May, 2024)

As Mistral’s top-tier model, Large excels in many multi-task language understanding tests, making it the second-best model (only slightly behind GPT-4) for complex language tasks. It also performed reasonably well in maths and code-generation tasks. It natively supports function calls and JSON output, making it easy to work with through APIs.

One key feature that stands out is its native fluency and cultural awareness in French, Spanish, German, and Italian, which gives it an edge in European market use cases.

Mistral offers a model self-hosting option for firms with strict data privacy requirements.

Mistral Moderation API

In November 2024, Mistral introduced the Mistral Moderation API [@mistralaiteamMistralModerationAPI2024] . This API can classify raw text or chat context in Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish into nine categories (with scores): sexual, hate and discrimination, violence and threats, dangerous and criminal content, self-harm, health, financial, law, and personally identifiable information.

Gen AI safety, modernisation, and grounding have been major headaches for many user-facing cases, especially for individual application developers and small companies that lack the resources to address these concerns. This API could be a welcome tool for simplifying Gen AI application development.

Mistral Batch API

Mistral released its Batch API on 7th November 2024 through its La Plateforme, offering up to 50% cost reduction compared to synchronous API calls. Batch API is ideal for use cases such as customer feedback and sentiment analysis, document summarisation and translation, creating embedding, and data labelling.

References