Mistral Small 3: Leading the Way in AI Model Efficiency

A New Contender in the AI Landscape

Mistral AI, a prominent name in the European artificial intelligence sector, has introduced an impressive new model: Mistral Small 3. With 24 billion parameters, it competes effectively with larger models like Llama 3.3 70B and Qwen 32B, particularly in the MMLU-Pro benchmark. Notably, it operates at higher speeds than the Llama 3.3 70B.

Surpassing OpenAI and Google Models

Mistral Small 3 stands out by offering superior performance compared to the widely-used GPT-4o mini on ChatGPT, a model often called upon when users exhaust their GPT-4o requests. Mistral Small 3 not only performs better but also benefits from reduced latency.

Open Source and Adaptability

Mistral AI has adopted an open approach for the Mistral Small 3, releasing both pretrained and instruction-tuned checkpoints under the Apache 2.0 license. This move encourages the open-source community to develop and customize this model further. Remarkably, the model hasn't been trained using reinforcement learning or synthetic data, differentiating it from others like Deepseek R1, which is a model it complements well. As a foundational model, Mistral Small 3 promises enhanced reasoning capabilities.

Accessibility and Versatility

One of the model's significant advantages is its accessibility. Those with high-performance personal computers can run Mistral Small 3 locally. It requires no more than a single Nvidia RTX 4090 graphics card or a MacBook with 32 GB of RAM.

Performance Insights

Despite its prowess in benchmarks, human evaluators did not always prefer it. When tested against over 1,000 proprietary coding and generalist prompts, Mistral Small 3 was favored over Gemma-2 27B and Qwen-32B but lagged behind Llama 3.3 70B and GPT-4o mini.

Availability

Currently, Mistral Small 3 is accessible on la Plateforme under the identifiers mistral-small-latest or mistral-small-2501.