What is Mistral Small 3? French rival to DeepSeek claims to outperform Qwen, ChatGPT and others as AI war intensifies | Mint

The battle of AI has taken the world by storm when DeepSeek became the topic of discussion across the globe. However, the latest move by Mistral AI with the release of Mistral Small 3 has added yet another layer of competition, offering a powerful open-source alternative to some of the leading proprietary AI models.

Mistral AI, a company known for its open-source artificial intelligence, has introduced Mistral Small 3, a latency-optimised model designed to rival significantly larger AI systems. With 24 billion parameters, the model is built to perform at a level comparable to Llama 3.3 70B and Qwen 32Bwhile outpacing them in speed, according to a Mistral AI blog post. Moreover, it provides an open replacement for proprietary models such as GPT-4o mini.

The company claims that Mistral Small 3 matches Llama 3.3 70B in terms of instruction-following capabilities while being more than three times faster when deployed on the same hardware. Its optimised architecture, designed with fewer layers than its competitors, allows for a significant reduction in processing time per query, making it a strong contender in the field of AI efficiency.

Benchmarks and accuracy

The Mistral Small 3 claims to have been pre-trained and instruction-tuned to cater to the majority of generative AI tasks. These include general language processing, instruction following, and fast-response applications. Unlike many competing models, it has not been trained using reinforcement learning or synthetic data, meaning it remains a flexible and customisable option for developers seeking to build upon its capabilities.

According to the company blog post, the model has demonstrated an impressive 81 per cent accuracy on the Massive Multitask Language Understanding (MMLU) benchmark while delivering a throughput of 150 tokens per second in internal benchmarking. 

To validate its effectiveness, Mistral AI engaged third-party evaluators to compare its performance against other models, including proprietary and open-weight alternatives. Their findings confirmed that Mistral Small 3 performs competitively in areas such as coding, mathematics, general knowledge, and instruction-following tasks.

Use cases and applications

The introduction of Mistral Small 3 opens up a range of possibilities across industries and applications. One of its key strengths is in fast-response conversational AI, where it can power virtual assistants requiring real-time interaction. The model’s efficiency also makes it a strong candidate for low-latency function calling, which is crucial for automated workflows and AI-driven decision-making.

Mistral Small 3 can be fine-tuned for specific domains, making it particularly valuable in industries such as healthcare, legal advisory services, and technical support. Companies seeking to deploy AI for specialised expertise can use this model as a foundation to build bespoke, domain-specific solutions.

Another notable feature is its ability to be deployed for local inference. As per Mistral AI, the model can be quantised to run efficiently on a single RTX 4090 GPU or a MacBook with 32GB of RAM, making it an attractive option for organisations prioritising data privacy and security. 

Open-Source AI in a competitive landscape

The AI space is evolving at a rapid pace, with companies vying to create the most effective and accessible models. The recent emergence of DeepSeek R1, another powerful open-source model, has demonstrated the increasing competitiveness of the sector. With Mistral Small 3 now in the mix, developers and businesses have a compelling new option that balances performance, efficiency, and accessibility.

For now, Mistral AI’s latest release has become another option for mid-sized AI models in an increasingly crowded landscape. Whether it will dominate this space remains to be seen, but one thing is certain—the battle of AI innovation is far from over.