Open source in AI: Meta and Mistral launch models challenging closed AI competitors

Last week, Meta released Llama 3.1 model, which features 405 billion parameters.

That's a massive model.

Until now, Meta had only released 8 billion parameter and 70 billion parameter Llama models. According to the company, the new model outdoes even the most advanced LLMs, such as ChatGPT 4-Omni and the Claude 3.5 Sonnet model from Anthropic, on many parameters. Zuckerberg also claims inferencing the Llama 3.1 costs a mere 50% of the cost of running ChatGPT 4o.

Compared to its previous version LLaMA 2, LLaMA 3 has better reasoning abilities, and code generation while also following human instructions effectively. It also outperforms other open models on benchmarks that measure language understanding and response (ARC, DROP and MMLU). All thanks to the revolutionary capabilities of LLaMA 3

On the other side of the Atlantic, Paris based Mistral released Large 2 model with 123 billion parameters.

The company claims the AI model offers significantly improved capabilities in code generation, mathematics, and reasoning. It also gets support for several new languages as well as advanced function calling capabilities.

Despite being one-third the size of Llama 3.1 405B, the company claims that its LLM outperforms it. Based on its internal benchmark testing, Mistral said its AI model fared better in code generation and math performance. It also claimed to outperform GPT-4o in Java code generation. Further, the company claims that the Mistral Large 2 has enhanced function calling and retrieval skills that allows it to power complex business applications. Function calling is a capability of AI models to interact with external tools or functions. This allows them to procure data from various sources and provide more accurate, informative, and efficient responses.

Notably, Mistral Large 2 is only available for research and non-commercial usages. Such limits are in contrast with the Open-source limit as the well as Meta’s approach of being fully open source.

Open-source advantages

Open-source software means the license holder allows the software's source code to be freely accessible and modified by outside parties. Outside developers can then make changes to the software, which allows them to improve functionality, fix bugs, or improve security.

By giving away the software code for free, developers have a chance to potentially improve the product quicker than a "closed" software system where only a company's employees have access to alter the code. As outside developers gravitate toward "free" open-source software rather than expensive proprietary software, the open-source model is best if one wants to scale up usage quickly. In addition to these general open-source advantages, Llama can run anywhere, so it doesn't force developers to send their private data to a closed model or specific cloud.

Mistral has partnered with Google Cloud Platform. It also available on cloud via Azure AI Studio, Amazon Bedrock, and IBM WatsonX.

Meta said its model will be available on AWS, Databricks, Google Cloud, Hugging Face, Kaggle, IBM WatsonX, Microsoft Azure, NVIDIA NIM, and Snowflake.

Monetization

A big factor in the AI race Meta's selection of open source for AI has the chance to make it the leading AI company in the world and, therefore, perhaps become the world's most valuable company one day. It will likely generate revenue by using the underlying model and building services on top of it. On the recent Q1 conference call with analysts, Zuckerberg hinted at building business messaging customer service AIs on WhatsApp, introducing ads into AI interactions in Meta AI, or perhaps charging for access to the largest AI workloads with more compute.

In the case of Mistral, the startup hopes customers will first experiment with its smaller open-source models and eventually move onto its larger, paid ones. The company plans to make money from these larger commercials models as well as its yet to be released APIs.

Other open-source software companies, such as Red Hat, have previously monetized AI models by selling customer service and consulting services. And WordPress has used a dual-license model to monetize its website-building software, offering large enterprises a "deluxe" paid version. So, there is precedent for profiting from open-source models.