Mistral’s Mixtral 8x7B AI Model: A Game-Changer in Open-Source AI
A New Era in Open-Source AI Models
Mistral, the trailblazer in open-source AI models and large language models (LLMs), has once again made waves with its groundbreaking release. The French startup, known for its dedication to open-source AI, recently unveiled its latest model called Mixtral 8x7B. This model employs the innovative “mixture of experts” technique, which combines different models specialized in various categories of tasks.
Last week, Mistral broke its tradition of grand announcements by unexpectedly dropping Mixtral 8x7B as a torrent link. With no accompanying blog post or demo video, the model’s capabilities were left to the imagination of early adopters and AI influencers on platforms like X and LinkedIn.
Benchmarking Success and Commercial Availability
Today, Mistral released a blog post shedding light on Mixtral 8x7B’s features and performance. The benchmarks showcased in the post demonstrate that the model rivals or even surpasses the closed-source GPT-3.5 by OpenAI and Meta’s Llama 2 family, the former leading the open-source AI race. Mistral also acknowledged the technical support provided by CoreWeave and Scaleway during the training process.
Furthermore, Mistral confirmed that Mixtral 8x7B is available for commercial use under an Apache 2.0 license. This announcement has spurred a wave of excitement among AI early adopters, who have eagerly downloaded the model and begun exploring its capabilities. Users have been astounded by its performance, especially considering its small footprint. The model can even run on machines without dedicated GPUs, such as Apple Mac computers featuring the new M2 Ultra CPU.
An intriguing aspect of Mistral 8x7B, as highlighted by AI influencer Ethan Mollick from the University of Pennsylvania Wharton School of Business, is its apparent lack of “safety guardrails.” This means that users who have felt restricted by OpenAI’s increasingly stringent content policies now have access to an equally powerful model capable of generating material deemed “unsafe” or NSFW by other models. However, this absence of safety measures may pose challenges for policymakers and regulators.
To experience Mixtral 8x7B for yourself, you can try it via HuggingFace, with credits to Merve Noyan for providing the link. It is worth mentioning that the HuggingFace implementation includes safety guardrails, as evidenced by its refusal to generate instructions on creating napalm when prompted.
Mistral’s future looks even more promising, with CEO Matt Schumer of HyperWrite AI revealing that the company has already launched an alpha version of Mistral-medium through its application programming interface (API). This development hints at the potential arrival of a larger and even more high-performing model in the near future.
In addition to the groundbreaking new model, Mistral has recently secured a massive $415 million Series A funding round, led by A16z. This funding has propelled the company’s valuation to an impressive $2 billion, further solidifying its position as the most successful startup in European history in terms of funding.
