Mistral’s NeMo rivals GPT-4o mini

Microsoft-backed, French AI start-up, Mistral, has launched its latest open-source “best small model”, called NeMo, which was built with strong backing from AI chipmaker, NVIDIA and is expected to rival OpenAI’s recently launched GPT-4o mini.

Like OpenAI’s GPT-4o mini, NeMo is designed for more affordable and cost-efficient high performance when it comes to tasks like reasoning, coding accuracy, and world knowledge.

On the industry’s standard benchmark test (MMLU) it scored a high 68% for its knowledge and problem-solving capabilities, which was a similar score to Google’s Gemma 2, but below OpenAI’s GPT-4o mini, which scored 82%.

Having been trained on over 100 languages, it does have impressive multilingual capabilities, and can accurately generate code and translate and summarize documents in languages such as  English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi.

And, because it’s open-source, it’s another step towards democratizing powerful language models, bringing “frontier AI models to everyone’s hands in all languages that form human culture.”

Mistral, which is just over a year old, was formed by ex-Meta and Google employees, successfully secured $645M in funding last month, and has formed a partnership with Microsoft, which gives Microsoft Azure cloud users access to its advanced LLMs.