When it rains, it pours, and if last week is anything to go by, it’s absolutely pouring AI launches. It all kicked off with Meta launching the largest open-source model yet, with a whopping 405 billion parameters. Next on the scene was OpenAI, which, out of nowhere, dropped SearchGPT—the company’s AI-powered alternative to Google.
Now Mistral, the French AI startup that recently raised $645 million, wants to steal the spotlight with its latest launch, Mistral Large 2, the company’s newest flagship AI model. The team insists that Large 2 is on par with and can surpass OpenAI’s GPT-4o and Meta’s Llama 405B when it comes to math, code generation, and reasoning.
To put the new model’s capabilities into perspective, Large 2 seemingly outpaces Llama 3.1 when it comes to code generation and math despite only shipping with a third of the parameters, 123 billion to be exact. Of course, AI benchmarks should be taken with a grain of salt.
According to the launch announcement, Mistral set their sights on a specific issue that has been plaguing AI models for months now: Hallucinations. You’re probably familiar with the phenomenon if you use any LLM regularly. Oftentimes, they can respond with totally made-up answers to even basic prompts. The company insists that Large 2 was trained to be more discerning in its responses, coming clean when it doesn’t know something rather than making something up.
Unlike its biggest AI competitors, Large 2 isn’t multimodal, meaning it can’t understand, analyze, and create information from different types of inputs like voice, video, and images. As it stands right now, Large 2 is text-only. If you’d like to test it out, it’s available now on Google’s Vertex AI, Amazon Bedrock, and more.