Mistral models, recognized for their open nature, have quickly become leaders among open-source LLMs. The company burst onto the scene by releasing capable open-weights models, Mistral 7B and Mixtral 8×7B. Mistral is partnering with Microsoft Azure to provide their models through Azure AI Studio and Azure Machine Learning in addition to their own platform and self-deployed options.
Mistral is perceived as a competitor to OpenAI, potentially altering leverage and negotiation dynamics within the industry by introducing more competition. This influences innovation, pricing, and access to AI technologies.
Their newest offering, Mistral Large, represents their most sophisticated language model to date. Now available through Mistral’s platform and on Azure, Mistral Large marks an exciting advancement in AI capabilities. Building on the success of Mistral Medium, which has achieved high rankings on the LMSys leaderboard, Mistral Large is expected to climb the leaderboards rapidly. Distinguished by exceptional performance in reasoning, knowledge benchmarks, multilingual capabilities, and tasks related to coding and math, Mistral Large showcases Mistral’s commitment to pushing the boundaries of what large language models can achieve. With strengths such as fluency in multiple languages, a 32,000 token context window, precise instruction following for content moderation, and native function calling, Mistral Large represents meaningful progress in conversational AI that can enable new applications and innovations.
Analysis
- Mistral appears to be stepping back from full ‘openness,’ at least for LLMs with abundant parameters. I’ve noted that the number of teams routinely publishing open LLMs is fairly limited. This quantity now seems diminished: Mistral’s choice to not plainly pledge open weights or models for Mistral Large, paired with its Microsoft affiliation, stokes worries about potential for more exclusive models moving forward. Although Google recently unveiled a duo of open models, Meta currently stands alone as the lone consistent source of open LLMs, at least for those over 30 billion parameters.
-
Despite the “open source” label on earlier models, a closer examination reveals Mistral LLMs don’t fully meet open source standards. They provide necessary components, such as model weights and appropriate licenses, for deployment and optimization but fall short in other open source criteria.
-
The practice of open sourcing models, while beneficial, introduces broader potential for security vulnerabilities.
- Comparing Mistral to OpenAI, the latter reported $2 billion in annualized revenue in December 2023, with projections to double this within a year. Despite this, OpenAI faces financial losses due to hefty expenses in research, development, personnel, and computing.
- OpenAI’s tools have demonstrated potential to bring value to companies in certain areas. As the capabilities develop further, businesses continue exploring optimal applications and use cases. For OpenAI’s offerings to reach their full potential, it is important that companies discover clear benefits in using them.
- Both OpenAI and Mistral face competition from other AI entities, which poses a threat to their market positions if they fail to maintain technological innovation or secure adequate funding.
- These AI startups are burning through traditional venture funding incredibly fast. Venture investors typically bet on early-stage companies where a large return on investment is possible. But for companies already valued at many billions of dollars, exponentially large outcomes are needed to generate those returns. So traditional VCs are largely priced out at this stage.
- As a result, AI startups like OpenAI, Mistral and Anthropic must court large companies and sovereign wealth funds that can write nine-figure checks. There are only a handful of such funders globally that can make a meaningful investment at this scale. Going public is also an option, but these startups have such massive capital needs that even public markets may not provide enough funding fuel for their ambitions.
- The enormous investment needs for pursuing Artificial General Intelligence (AGI) ambitions may necessitate these startups to reconsider their goals. While the allure of AGI fits OpenAI’s ambitious vision, the practical reality is that most companies desire AI solutions for specific business needs rather than a computational cure-all. The path forward lies in developing profitable products that solve these real-world problems, even if less glorious than AGI.
The pace of innovation in AI and foundation models brings many benefits. Developers and AI teams now have numerous model options, with impressive new foundation models regularly released. As I recently highlighted, I’m particularly excited about Google’s Gemini family of models. Their capabilities push boundaries while maintaining rigor around ethics and transparency.
If you enjoyed this post please support our work by encouraging your friends and colleagues to subscribe to our newsletter:
