French AI corporation Mistral has unveiled Europe's first reasoning AI model, entering the global AI competition that is dominated by the United States and China. As reasoning AI models, capable of independent thought and judgment, emerge as the core of the AI competition, latecomer countries are also joining the fray. However, with OpenAI and Google releasing upgraded versions, the competition for reasoning AI models is expected to intensify further.
According to foreign media outlets such as Reuters and CNBC on the 12th, Mistral unveiled its reasoning AI model 'Magistral' on the 10th (local time). 'Magistral' is an advanced model equipped with the ability to solve complex problems step by step, excelling in tasks that require logical thinking, such as mathematical calculations and coding. Notably, it features an intermediate reasoning process and is available in two versions: 'Magistral Small' offered as open source and 'Magistral Medium' targeted at corporate customers.
Mistral stated that 'Magistral' offers faster response speeds compared to competitive models. By utilizing the new thinking mode and flash-response feature of its AI chatbot 'Le Chat', users can receive answers ten times faster than competitors. However, 'Magistral Medium' scored lower than Google's 'Gemini 2.5 Pro' and Anthropic's 'Claude Opus 4' in physical, mathematical, and scientific tests. It also did not surpass 'Gemini 2.5 Pro' in programming benchmarks.
The industry is watching closely to see if Europe will stand out in the AI competition, which has been led by the United States and China, following the unveiling of the continent's first reasoning AI model.
The development of reasoning models is also active in Korea. LG AI Research unveiled the country's first reasoning model 'Exawon Deep' in March. Naver Cloud is about to launch a reasoning model developed based on its flagship model 'HyperCLOVA X'. Korean AI startups are also busy. Last month, Konan Technology announced the reasoning-non-reasoning integrated model 'Konan LLM Enterprise (ENT)-11', while Saltlux unveiled the 'Lucia3' model that includes reasoning capabilities.
Reasoning models are emerging as a key component in the AI competition as they are seen as a steppingstone toward artificial general intelligence (AGI). AGI refers to AI that surpasses human intelligence levels, possessing the ability to autonomously analyze new problems and find solutions without relying on training data. OpenAI has previously stated that reasoning models are categorized as the second stage in its five-step roadmap toward AGI.
In the past, reasoning models required thousands of high-performance graphics processing units (GPUs), each costing millions of won, along with considerable data expenses, power, and labor costs, creating high barriers to entry. However, it is interpreted that AI latecomer countries have become able to launch reasoning models due to increased cost efficiency. By focusing on 'reinforcement learning', where AI learns by itself during the reasoning process through open-source frameworks, it has become possible to develop low-cost, high-performance AI without expanding model size or computational resources. Notably, Chinese AI startup DeepSeek has unveiled R1, changing the landscape for reasoning models.
Competitors that have early launched reasoning models are further enhancing their technological capabilities. On the day 'Magistral' was unveiled, OpenAI released 'o3-Pro', an upgraded version of its existing reasoning model. This model is a higher version of the reasoning model o3 that OpenAI introduced earlier this year. o3-Pro, which focuses on performance, is convenient for tasks such as web search, file analysis, visual input reasoning, Python usage, and response customization using memory. Google also revealed the advanced reasoning feature 'Deep Think' for 'Gemini 2.5 Pro' at last month's developer conference. 'Deep Think' enhances the AI's cognitive abilities by considering multiple possibilities in parallel rather than providing a single correct answer.