Mistral AI drops its new model with mixture of experts.

Mistral AI has released Mixtral 8x7B, an open-source sparse mixture-of-experts model with 45B parameters that achieves state-of-the-art performance among open models. Mistral models are now available via its dev platform and API in beta.

What's going on here?

French AI startup Mistral releases a new open-source model.

What does this mean?

Mixtral matches or exceeds LLaMA 2 70B and GPT-3.5 on most NLP benchmarks, with 6x faster inference speed than LLaMA models. Mixtral uses a “mixture of experts” network that selectively uses only 12B of its 45B total parameters per inference, keeping costs similar to a 12B model. [Mostly everyone believes] that GPT-4 is also a mixture of experts.

Relevant links: magnet for the model [1], Mixtral on Replicate [2], Mixtral fine-tuned for chat [3].

Mistral now also has API endpoints and a platform for these models. There are 3 generative endpoints (Mistral-tiny, small and medium) and one embeddings endpoint (Mistral-embed). Mistral medium runs on another unreleased model with higher performance.

Why should I care?

Just like their first model, Mistral realised this one by tweeting a magnet to download the model. But another factor that boosted Mistral’s launch was Google’s edited demo of Gemini without any access to the models.

Mistral also raised at a $2B valuation just a few days back. Reminder: the startup was formed this year in June. It’s releasing the best-in-the-class open-source models for everyone to use. Along with Llama, these models are gearing to take OpenAI’s market share.