Mistral unveils its new opponent to the DeepSeek R1 and OpenAI o3 models
As competition in the field of reasoning models intensifies, Mistral AI, a fast-rising French startup, presents its adaptation to the challenges posed by players like DeepSeek and OpenAI. With the launch of its new range of tools, including the Magistral model, Mistral aims to establish itself in a market where performance and innovation have become crucial.
Mistral AI aims to compete with models such as DeepSeek R1 and OpenAI o3 by offering solutions adapted to the diverse needs of developers and businesses. This approach is part of a rapidly evolving technological context, where the speed and efficiency of reasoning model responses can make all the difference. Let’s dive into this promising innovation and discover what sets Mistral apart from its competitors.
Introducing Magistral: A New Reasoning Model from Mistral AI
The Magistral model, recently launched by Mistral AI, comes in two variants: Magistral Small and Magistral Medium. The first, available under an open Apache 2.0 license, is designed with 24 billion parameters, while the second is a proprietary version. These models are based on a robust architecture, built on the Mistral Small and Medium 3.1 platform, providing a competitive base in the market. Unlike other companies that rely on pre-existing data, Mistral AI has opted for a unique approach. The startup has developed its own learning pipeline, using the Reinforcement Learning from Verifiable Rewards (RLVR) technique. This strategic choice demonstrates Mistral’s commitment to delivering high-performance models adapted to modern requirements. Adopting an innovative approach to learning Mistral AI’s method for training its models is based on principles from policy optimization, while avoiding the use of common techniques such as Proximal Policy Optimization. Instead, the startup relies on Group Relative Policy Optimization (GRPO), inspired by the work of DeepSeek. By combining rewards from multiple outcomes, Mistral AI develops models whose learning refines over time, particularly in fields such as mathematics and programming. To enhance the user experience, Mistral AI has also taken care to adapt its models to multiple languages. By translating 10% of its English-language problems into languages such as French, Spanish, Italian, German, Chinese, and Russian, the company avoids the problems of language mixing, a problem reported by DeepSeek users. The performance improvement is undeniable. On AIME scientific benchmarks, Mistral’s models’ scores show results ranging from 4.3% to 9.9% higher in English than in other languages. This focus on linguistic diversity demonstrates Mistral AI’s ambition to remain competitive in a global market.Architecture and Training Samples Mistral’s training process also focused on optimizing training data. From 700,000 mathematical samples, the startup used a system of rules to select approximately 38,000 problems and solutions to create a refined version of Mistral Large 2. Similarly, for programming, 35,000 problems were integrated into Magistral Medium’s training.This vast database strengthens the models’ learning capacity. As Magistral Medium improves its performance, the complexity of the problems submitted also increases. This ensures that the models don’t just learn simple answers, but can also handle complex and varied scenarios.
Features Magistral Small Magistral Medium
Number of Parameters
24 billion Proprietary License
Apache 2.0 Proprietary Pipeline Optimization
Group Relative Policy Optimization Group Relative Policy Optimization Training Data
38,000 Math Problems
38,000 Problems + 35,000 Code Problems To give users an idea of the models’ performance, Magistral AI has chosen to frame some of its training phases with supervised tests that allow for adapting complexity levels. By mixing the prompts, the company found that this diversity is fundamental to a good start to reasoning.Evaluating Performance Against the Competition Despite its innovations and unique methodology, Mistral AI faces a significant challenge: fierce competition from giants like DeepSeek and OpenAI. Magistral Medium’s results perform well, but remain below those of the market leaders. It is reported that under similar evaluation conditions, this model achieves results close to those of DeepSeek’s R1 Zero and R1 models, without however surpassing them. For Mistral, the goal is not only to compete, but to constantly refine its approach. Mistral researchers point out that, although Magistral Small’s performance stops increasing after 40,000 tokens, the theoretical context window reaches 128,000 tokens. This ranking remains a crucial point in the startup’s career, as it aims to transcend these limitations in its future iterations. Technology and Supreme Speed
One of the remarkable features of Mistral’s models is their speed of execution.
| Magistral Small | can run on a single RTX 4090 GPU with 24 GB of VRAM, contrasting with the higher hardware requirements of some other models. For users on various platforms, an optimized version of this model is also in preparation for Apple Silicon computers. | In terms of responsiveness, Mistral claims that its Magistral Medium solution can generate answers up to 10 times faster than its direct competitors. For example, while OpenAI o3 takes approximately 40 seconds to provide an answer, Mistral Medium can do so in just 10 seconds. However, it is important to note that depth of reasoning remains a key criterion. Accessibility and Market Integration |
|---|---|---|
| The availability of Mistral AI models on platforms such as Amazon SageMaker, IBM Watsonx, and Azure AI marks a significant step toward the startup’s ambition to capture significant market share. Users can also consider deploying these models on-premises by contacting Mistral AI’s sales team, facilitating custom integration. | Mistral’s promise to provide businesses with a trail of every response generated by its models helps reassure customers of the transparency and quality of results. These elements are crucial in sectors such as research, data analysis, and informed decision-making. Criteria | Magistral Small |
| Magistral Medium | Response Time | 10 seconds |
| 4 times faster than OpenAI o3 | GPU Required | 1 x RTX 4090 |
| 1 x RTX 4090 | Access to Platforms | Hugging Face, The Platform |
Preview on The Platform, Amazon SageMaker
To conclude this section, Mistral AI doesn’t just create competitive models, but also innovates to redefine market standards. With a strong commitment to performance and rapid response, the French startup envisions a future where it aspires to compete with the best in the field.
Vision for the Future: Beyond Prototypes As Mistral AI prepares for its ramp-up, the startup’s vision for the future seems clear. The goal is to continue innovating and developing models capable of elevating the capabilities of artificial intelligence systems to new heights. To this end, the company plans to focus on the following aspects: Regular Iterations:Mistral AI is committed to constantly improving its models to ensure they remain at the cutting edge of technology.
Strengthening Multimodal Capabilities: Although Magistral Medium and Small were trained solely on text data, integrating multimodal capabilities remains an area of future innovation. Adapting to user needs: Understanding business expectations and adapting models accordingly will remain a priority. Expanding accessibility:Ensuring solutions are available to a wider audience, from independent developers to large enterprises.
Mistral also plans to implement tools that improve reasoning speed while preserving the quality of responses. By capturing user feedback, the startup intends to further adjust its offering, using real-world data to refine its algorithms.
The ability to evolve in such a dynamic sector is essential. As technological innovation continues to advance, companies like Mistral AI must remain agile to evolve with their customers’ needs and stand out in an increasingly saturated market. Expert Support and Strategic Partnerships To strengthen its capabilities, Mistral AI is considering strategic partnerships with research institutions and other technology companies. By collaborating with industry experts, Mistral hopes to leverage the latest scientific advances to propel its models to new horizons. By fostering a collaborative ecosystem, the startup can not only benefit from external expertise but also position itself as a key player in the sustainable innovation of reasoning technologies.In short, as Mistral AI prepares to take on established rivals like DeepSeek and OpenAI, the startup’s future aspirations rest on a solid foundation of performance, innovation, and collaboration. The coming months will be crucial to observe how Mistral will manage to exceed expectations and transform challenges into opportunities.