Breaking News: Mistral AI Takes on OpenAI, Google, and Meta with Mixtral 8x7B
This week, Mistral AI, a Paris-based AI startup, stole the spotlight from Gemini, releasing its latest model, Mixtral 8x7B. The buzz around Gemini quickly faded as Mistral AI showcased its new creation. The startup, which raised over $113 million in June, has positioned itself as a serious contender in the AI space.
Key Features of Mixtral 8x7B
- Architecture: Based on the Mixture of Experts (MoE) architecture.
- Licensing: Licensed under Apache 2.0.
- Performance: Outperforms Llama 2 70B on most benchmarks, boasting 6x faster inference.
Clever Branding and Funding Success
Mistral AI cleverly brands itself as the 'Mixtral of Experts,' a play on the training strategy employed by OpenAI for GPT-4. Mixture of Experts allows models to be pre-trained with significantly less computational power, enabling scalability without breaking the budget.
Mixtral 8x7B, a decoder-only model, efficiently manages its parameters using a router network. This network selects two expert groups to process each token, enhancing the model's performance. Despite having a total of 46.7 billion parameters, Mixtral effectively uses only 12.9 billion parameters per token, achieving comparable speed and cost efficiency to a 12.9 billion-parameter model.
Mistral AI is on a funding spree, securing $415 million with a $2 billion valuation. The funding, led by Andreessen Horowitz and Lightspeed Venture Partners, positions Mistral AI for further growth.
In a strategic move, Mistral AI introduced 'La Plateforme,' offering API endpoints for its models. The models are categorized into Mistral Tiny, Mistral Small, and upcoming Mistral Medium. Notably, Mistral Medium aims to outperform models like GPT 3.5, supporting multiple languages and achieving an impressive score of 8.6 on MT-Bench.
Unlike some open-source competitors, Mistral AI opted to launch a paid endpoint for Mistral Medium. This approach allows them to gather customer feedback, iterate on real-world use cases, and monetize their models effectively.
Mistral AI's decision to remain exempt from the EU AI Act, along with other open-source companies, signals a strategic move to protect trade secrets. As the European Union sets rules for AI, Mistral AI's unique position allows flexibility in navigating regulatory requirements.
While Mistral AI started as an open-source company, speculation arises about future models not being released as open source. This echoes OpenAI's trajectory, which shifted its stance on open sourcing as it evolved.
Conclusion: A Rising Star in the AI Landscape
In a market dominated by giants like OpenAI, Google, and Meta, Mistral AI's rise with Mixtral 8x7B showcases its commitment to making generative AI exciting and competitive. The startup, with top-notch marketing and promising products, aims to establish itself as a formidable player in the AI landscape.