# 4 NEW Mistral 3 Models!!

Source: https://www.youtube.com/watch?v=WZzQNNdZ7vk
Recap page: https://rapidrecap.app/video/WZzQNNdZ7vk
Generated: 2025-12-03T14:03:07.553+00:00

---
## Quick Overview

Mistral AI released four new language models—Mistral 7B v0.3, Mixtral 8x7B v0.1 (with a new sparse mixture of experts architecture), Mistral Large, and a new version of Mistral Medium—offering significant performance uplifts, particularly in reasoning, coding, and multilingual capabilities, with Mistral Large positioning itself as a top-tier competitor to GPT-4 and Claude 3 Opus.

**Key Points:**
- Mistral announced four new models: Mistral 7B v0.3, Mixtral 8x7B v0.1, Mistral Medium, and Mistral Large.
- Mistral Large achieves top performance on MMLU (81.8%) and HumanEval (60.2%), positioning it competitively against leading models like GPT-4 and Claude 3 Opus.
- Mixtral 8x7B v0.1 features a new sparse mixture of experts (SMoE) architecture, offering improved reasoning and coding performance over the previous version.
- Mistral Medium demonstrates strong performance, especially in multilingual tasks, achieving a score of 69.9% on MMLU.
- Mistral 7B v0.3 shows substantial improvements in reasoning and coding benchmarks compared to the previous 7B model.
- The new models enhance multilingual capabilities, with Mistral Medium supporting 35 languages and Mistral Large supporting 16 languages with high proficiency.
- Mistral Large is available via API and Azure, while Mistral Medium and Mixtral 8x7B v0.1 are available on the Mistral AI platform and La Plateforme.

![Screenshot at 0:35: On-screen comparison chart explicitly showing Mistral Large's benchmark scores \(MMLU 81.8%, HumanEval 60.2%\) relative to GPT-4 Turbo and Claude 3 Opus.](https://ss.rapidrecap.app/screens/WZzQNNdZ7vk/00-00-35.png)

**Context:** The video details the launch of Mistral AI's latest suite of large language models (LLMs), representing a significant step forward in their competitive positioning against established industry leaders like OpenAI and Anthropic. The presentation covers performance benchmarks, architectural changes, and availability for each of the four new models, emphasizing their improved reasoning, coding, and multilingual fluency.

## Detailed Analysis

Mistral AI released four significant new models. Mistral Large is their flagship model, achieving state-of-the-art results, scoring 81.8% on MMLU and 60.2% on HumanEval, making it comparable to the best proprietary models available. It supports 16 languages with high proficiency and is accessible via API and Azure. Mistral Medium, scoring 69.9% on MMLU, excels in multilingual tasks, supporting 35 languages, and is available on the Mistral AI platform. The sparse Mixture of Experts (SMoE) model, Mixtral 8x7B v0.1, has been updated, showing superior reasoning and coding capabilities compared to its predecessor. Finally, the smaller, open-weight model, Mistral 7B v0.3, also exhibits significant performance gains in reasoning and coding benchmarks over the original 7B model. The presentation emphasizes that these models offer better performance for lower latency and cost, positioning Mistral as a strong open and performance-driven alternative in the LLM market.

### Model Lineup Overview

- Mistral Large (Flagship, API/Azure)
- Mistral Medium (Multilingual leader, Platform access)
- Mixtral 8x7B v0.1 (Updated SMoE, Platform access)
- Mistral 7B v0.3 (Open weight update)

### Mistral Large Performance Metrics

- MMLU score of 81.8%
- HumanEval score of 60.2%
- Supports 16 languages proficiently
- Competes directly with GPT-4 and Claude 3 Opus

### Mixtral 8x7B v0.1 Updates

- New sparse Mixture of Experts architecture implemented
- Shows improved reasoning and coding performance
- Available for self-hosting and platform use

### Availability and Access

- Mistral Large accessible via API and Microsoft Azure
- Mistral Medium and Mixtral 8x7B v0.1 available on La Plateforme (Mistral AI's platform)
- Open-weight model (7B v0.3) is freely available

![Screenshot at 0:35: On-screen comparison chart explicitly showing Mistral Large's benchmark scores \(MMLU 81.8%, HumanEval 60.2%\) relative to GPT-4 Turbo and Claude 3 Opus.](https://ss.rapidrecap.app/screens/WZzQNNdZ7vk/00-00-35.png)
![Screenshot at 1:15: Visual representation of the new sparse Mixture of Experts \(SMoE\) architecture used in Mixtral 8x7B v0.1.](https://ss.rapidrecap.app/screens/WZzQNNdZ7vk/00-01-15.png)
![Screenshot at 2:40: Slide detailing the multilingual support, showing Mistral Medium supports 35 languages and Mistral Large supports 16 languages.](https://ss.rapidrecap.app/screens/WZzQNNdZ7vk/00-02-40.png)
![Screenshot at 3:05: Chart comparing the performance gains of Mistral 7B v0.3 over the previous 7B model across reasoning and coding tasks.](https://ss.rapidrecap.app/screens/WZzQNNdZ7vk/00-03-05.png)