Mistral AI is holding talks to raise about $600 million from investors at a $6 billion valuation, the Wall Street Journal reported today.

The news comes less than a month after sources told The Information that the artificial intelligence model company is seeking new funding at a $5 billion valuation. The disparity between the two sums may indicate that investor interest in Mistral’s round has gone up since The Information’s report. According to the Journal, returning investors General Catalyst and Lightspeed Venture Partners are expected to be among the biggest contributors to the raise. 

Paris-based Mistral launched last April and raised about $500 million over the following eight months. The company’s most recent funding round, a $415 million investment announced in December, valued it at $2 billion. 

That Mistral’s potential funding round is expected to triple its valuation hints investors have found additional reasons to be optimistic about its growth prospects. The company, which launched its first paid products in February, may be experiencing strong revenue momentum. Alternatively, the expected valuation jump might reflect yet-undisclosed technical milestones in Mistral’s product development efforts.

The company has so far released a trio of open-source neural networks headlined by Mixtral 8x22B, a large language model that debuted last month. The LLM correctly answered 77.75% of the questions in MMLU, a popular benchmark test for comparing AI models. That’s just under the 79.5% score achieved by the most advanced edition of Meta Platforms Inc.’s new Llama 3 LLM.

The Llama 3 edition in question includes 70 billion parameters. Mixtral 8x22B has about twice as many, but activates only 39 billion parameters to generate prompt responses. That significantly reduces the LLM’s hardware usage, which in turn lowers inference costs. 

Mixtral 8x22B’s efficiency stems from the fact that it features a so-called “mixture of experts” architecture. LLMs based on this design comprise multiple, smaller neural networks that are each optimized for a different set of tasks. When it receives a prompt, Mixtral 8x22B activates only the neural networks that are best-equipped to generate an answer and keeps the rest dormant, which lowers hardware requirements.

Mistral also offers a collection of paid, cloud-based LLMs headlined by Mistral Large. That model includes a so-called function calling feature that allows it to perform tasks in other applications. Developers also have an option to package Mistral Large’s output into the JSON file format, which makes it easier to make AI responses available through a company’s custom software.

It’s possible the new funding Mistral is seeking to raise will go toward the development of additional, more capable LLMs. If rival OpenAI’s product strategy is any indication, it’s also possible the company will introduce other types of models besides LLMs for tasks such as video generation. 

Mistral has already started expanding beyond the language model segment. One of its paid offerings, Mistral Embed, is an AI designed to turn text into embeddings, mathematical structures that are easier for neural networks to process than raw data. The model is positioned as a more capable alternative to fastText, one of the most popular open-source tools for creating embeddings.

Photo: Mistral