JM Robles - Tech Consultant & Entrepreneur

Mistral: The European Alternative to ChatGPT

Mistral, a startup based in Paris, has become a European alternative to ChatGPT. The company recently raised a funding round of 500 million euros, elevating it to unicorn status.

About Mistral

Mistral has focused on developing artificial intelligence and machine learning technology, and their flagship product is the latest generation language model they have recently launched. You can learn more about Mistral on their official page.

Mistral’s Model

The same day that Google introduced us to its Gemini model, or rather a doctored video of it, the Mistral team released their new model via BitTorrent.

The model proposed by Mistral is a mix of experts, also known as MoE model (Mixture of Experts). It consists of having 8 models with 7B parameters each. Each of these 8 models has been trained to have better behavior in a specific quality. When generating tokens, based on the prompt, the model decides which submodel or submodels should generate the token. Apparently, GPT-4 has a similar approach.

What are MoE Models?

MoE models are a machine learning technique that combines the predictions of several models (the “experts”) to produce a final prediction. The idea is that different models may be good at different parts of the input space, and combining their predictions can lead to a more accurate and efficient model.

Model Performance

In benchmarks, Mistral’s model seems to outperform both GPT 3.5 and Llama 70B, demonstrating the effectiveness and efficiency of their approach.

Access to the Model

In addition to being able to access the model via BitTorrent, it is also possible to use it directly on Hugging Face.

But that’s not all, Mistral has launched its own API service to access the model in a similar way to how OpenAI does.

This is undoubtedly great news for everyone. Competition is good, and even better if it’s made in Europe. Personally, I have already requested access to the API (which is currently in whitelist mode). As soon as I have the chance to try it, I will share my impressions.

Mistral

See also