French generative artificial intelligence startup Mistral AI announced today it has closed on its second bumper funding round of the year, raising 385 million euros, or around $415 million, in a deal led by prominent venture capital firms Andreessen Horowitz and Lightspeed Venture Partners.

The latest round is said to bring Mistral AI’s value to around $2 billion, and is yet another sign that suggests investors are unlikely to quench their thirst for promising generative AI startups anytime soon.

The Paris-based startup’s value has risen more than sevenfold just six months after bursting onto the scene in June, when it raised $113 million in a seed funding round just four weeks after it was founded. Its three co-founders all have significant pedigree in the AI industry. Timothée Lacroix and Guillaume Lample both previously worked as researchers at Meta Platforms Inc’s Paris AI Lab, while Arthur Mensch spent time working at DeepMind, an AI research lab that has been a part of Google LLC since 2014.

Like its competitors OpenAI and Google, Mistral AI is a developer of large language models, which are a kind of AI that powers chatbot technologies such as ChatGPT and Gemini. ChatGPT famously took the internet by storm late last year with its humanlike ability to hold conversations with users on almost any topic.

But Mistral AI differs in its approach. It’s a firm believer that generative AI technologies should be open-source, meaning that the code that underpins its LLMs is freely available for others to use and modify. By taking this open-source approach, the startup is aiming to give other users the tools to quickly build their own, customized chatbots.

OpenAI and Google have said in the past that it’s dangerous to do things this way, arguing that the underlying LLMs can be abused and used to create tools that spread disinformation and other malicious content.

“We just believe A.I. should be open,” Anjney Midha, a partner at Andreessen Horowitz, told the New York Times in an interview. He argues that the open-source approach has emerged as the norm in just about all other areas of technology, including computer operating systems, programming languages, databases and more.

OpenAI, Google and Microsoft Corp. are currently believed to be leading the race in generative AI development, having spent billions of dollars to train models such as GPT-4 and Gemini. These models are trained using enormous amounts of text and data from the internet, and have the ability to answer questions, write exams and poetry, create original computer code and more.

Some argue that it’s incredibly dangerous to release such powerful models in the open-source domain. OpenAI and Google have notably spent months developing safety guardrails for their LLMs to ensure they cannot be used to spread disinformation and hate speech, for example, or generate biased answers to questions.

However, others believe that the generative AI crown will ultimately be won by those companies that give the technology away for free, even if that means taking the risk of open-sourcing it without many of the same guardrails.

This is one of the deeper discussions about AI development, analyst Holger Mueller of Constellation Research Inc. told SiliconANGLE. “Microsoft, Google and OpenAI all believe AI development should be closed, whereas the likes of Mistral AI believe it should be open source,” Mueller explained. “There are pros and cons to both approaches, but in most other aspects of the technology industry, such as infrastructure-as-a-service and platform-as-a-service, open source has emerged as the clear winner.”

However, that is not always the case, because in the area of software-as-a-service, most of the leading platforms remain proprietary software. “So we can’t be sure which approach will win for AI,” Mueller added.

The leading developer of open-source LLMs at the moment may be Meta, which released its popular Llama 2 model earlier this year. Anyone can take the foundational Llama 2 model and fine-tune it with their own data to create chatbots that are customized to perform different tasks.

Mistral AI’s product is similar. In September, it announced the availability of Mistral 7B, its first-ever LLM. It’s a relatively small-sized LLM with just 7 billion parameters, but the startup claims it can outperform many of its larger rivals by processing and generating responses more efficiently and at lower costs.

Mistral AI’s ambitious roadmap calls for the development of additional “frontier models” focused on performing tasks such as summarization and responding to questions.

Mensch, who is Mistral AI’s chief executive officer, says the company has designed a more efficient and cost-effective training method for its LLMs. According to him, its models can operate at less than half the cost of either OpenAI’s or Google’s best LLMs.

Image: Mistral AI