Mistral Astounds AI Community as New Open Source Model Surpasses GPT-3.5 Performance

Mistral, the most well-funded startup in European history, is a French company focused on open-source AI models and large language models (LLMs). Recently, it made waves in the AI community with the release of its new model — Mixtral 8x7B. This innovative model utilizes a "mixture of experts" approach, combining various specialized models to excel in different tasks.

In a distinctive, low-key fashion, Mistral released Mixtral 8x7B online as a torrent link, without any accompanying explanation, blog post, or demo video. This approach drew immediate attention from early adopters and AI influencers on platforms like X and LinkedIn.

Today, Mistral published a blog post detailing Mixtral 8x7B’s performance benchmarks, where it matches or even surpasses OpenAI’s proprietary GPT-3.5 and Meta’s Llama 2, the former leader in open-source AI. The company revealed its collaboration with CoreWeave and Scaleway for technical support during the model’s training and confirmed that Mixtral 8x7B is available for commercial use under an Apache 2.0 license.

Early adopters have already downloaded Mixtral 8x7B, and many are impressed with its performance. Its compact design allows it to run locally on standard machines, including Apple Mac computers equipped with the new M2 Ultra CPU.

Notably, Ethan Mollick, a professor at the University of Pennsylvania’s Wharton School and an AI influencer, highlighted on X that Mixtral 8x7B appears to have "no safety guardrails." This characteristic could appeal to users frustrated with OpenAI’s stringent content policies, enabling them to produce content often labeled as “unsafe” or NSFW by other models. However, this lack of guardrails poses potential challenges for policymakers and regulators.

You can explore Mixtral 8x7B yourself via HuggingFace (thanks to Merve Noyan for the link). The HuggingFace implementation does incorporate guardrails, demonstrated when testing it with controversial prompts — it declined to provide instructions for creating napalm.

Furthermore, Mistral is already developing even more powerful models. Matt Schumer, CEO of HyperWrite AI, mentioned on X that the company has launched an alpha version of Mistral-medium on its application programming interface (API), indicating that a larger, more advanced model is on the horizon.

In a significant financial boost, Mistral recently closed a $415 million Series A funding round led by A16z, achieving a valuation of $2 billion.