A new model approaching GPT-4 is released! “European version of OpenAI” teamed up with Microsoft, but was questioned as violating its original intention

Last night, Mistral AI, known as "European OpenAI", released its latest top-level text generation model, Mistral Large.

This model has top-level reasoning capabilities and can be used to handle complex multi-language reasoning tasks, covering text understanding, transformation and code generation.

Simply draw the key points:

  • 32K context window to accurately extract large document information.
  • Precise command following capabilities allow developers to customize audit strategies.
  • Supports function call and output mode restrictions to help scale application development and modernize technology stacks.
  • Native support for English, French, Spanish, German and Italian, with deep understanding of grammar and cultural background.

Attached is the experience address: https://chat.mistral.ai/chat

Mistral AI expands its skills again

Mistral Large has performed well in multiple benchmark tests, becoming the second most widely used model in the world through API, second only to GPT-4, and leaving other mainstream models behind.

Compared with LLaMA 2 70B, GPT-4, Claude 2, Gemini Pro 1.0 and other mainstream models, Mistral Large shows strong strength in reasoning capabilities.

In multiple common sense and reasoning benchmarks such as MMLU, Hellas, and WinoG, Mistral Large follows GPT-4 and far exceeds other models.

Mistral Large performs significantly better than LLaMA 2 70B on the HellaSwag, Arc Challenge, and MMLU benchmarks in French, German, Spanish, and Italian.

Mistral Large also performed well in coding and math tasks. Many benchmark tests are still far ahead.

In addition, Mistral AI also released a new model optimized for latency and cost, Mistral Small, which surpasses Mixtral 8x7B in performance but has lower latency.

In Mistral AI's view, Mistral Small will be an ideal middle choice between open source and flagship models, enjoying the same innovative advantages as Mistral Large, especially in terms of RAG (Retrieval Augmentation Generation) enablement and function calls.

However, the low-latency characteristics claimed by the Mistral Small model are limited compared to the Mixtral 8x7B, but its cost is significantly increased, with the input cost being 2.8 times higher and the output cost being as high as 8.5 times.

Mistral Large is now available on La Plateforme and Azure platforms, and integrated into Le Chat. Le Chat (literally translated as cat in French) is an AI chatbot similar to ChatGPT that supports multi-language communication.

Users can choose between three different models: Mistral Small, Mistral Large and Mistral Next. However, Le Chat does not support real-time networking functionality.

Shortly after launch, Le Chat crashed due to a larger-than-expected crowd flocking to the experience, and as of press time, Le Chat is back up and running.

We also experienced Le Chat for the first time. Depending on the complexity of the language itself and the amount and quality of training data it receives, Le Chat is more expressive in English and may vary from language to language in other languages.

Faced with the Chinese brainteaser, Mistral Large was obviously a little confused, and he didn't give an answer for a while.

Question: Butterfly, ant, spider, centipede, they work together, which one does not get paid in the end?

But if you switch it to English, it immediately transforms into a problem-solving expert. The answers are clear and logical, and it rises to a "human-centered perspective of work and rewards." However, the brain teasers in the Chinese context do not follow common sense. After all, "Centipede" is not rewarded.

The problem of chickens and rabbits living in the same cage is not a piece of cake for Mistral Large. After posing the linear equations, the answer came out quickly, but the 16 chickens and 19 rabbits in the answer were incorrect.

Question: There are chickens and rabbits in the same cage, with a total of 35 heads and 94 legs. How many are the chickens and rabbits?

This is a classic mathematical problem that can be solved using a system of linear equations.
Assuming that the number of chickens is x and the number of rabbits is y, the following two equations can be listed:
1. x + y = 35 (the total number of chickens and rabbits is 35)
2. 2x + 4y = 94 (the chicken has 2 legs and the rabbit has 4 legs, so there are 94 legs in total)
Express x in the first equation as a function of y to get x = 35 – y, and substitute it into the second equation to get:
2(35 – y) + 4y = 94
Solving the equation gives us y = 19. Substituting in x = 35 – y gives us x = 16.
Therefore, there are 16 chickens and 19 rabbits.
The latter two questions can be solved in the same way. The answers are 16 chickens and 19 rabbits.

Then we asked ChatGPT to give Mistral Large a math test question, which was more intense, but the overall performance was good:

Small model, big achievement

Mistral AI was founded in May last year with only 25 employees, but its large language model has already become famous on Hugging Face.

Previously, we reported that it released Mixtral 8x7B. Once this model was launched, it instantly detonated the open source community and kept countless developers awake at night.

Although it only has a parameter size of 46.7B, the performance of Mixtral 8x7B, which is good at small and large, is comparable to Llama 2 70B and GPT-3.5 in many benchmark tests, and even slightly better in some tests.

The secret of Mistral AI’s success lies in its clever integration of the three core elements in the AI ​​field—talent, data, and computing power.

The Economist revealed that Mistral AI’s founders and technical backbones all came from France’s elite technical institutes and have accumulated valuable industry experience in the research laboratories of technology giants such as Google and Meta. They are one of the few experts in the world who truly master how to train cutting-edge model technology.

Secondly, data is another magic weapon for Mistral AI’s success. Mistral is unique in model data training. For example, the model can effectively filter out repetitive or meaningless information, making the model more streamlined and efficient, with a parameter scale of only billions.

This means that ordinary users can even easily run Mistral AI models on their own personal computers.

For the arrival of Mistral Large, Turing Award winner Yann LeCun, NVIDIA senior scientist Jim Fan and other big names also sent congratulations on X.

At present, Mistral AI's valuation has exceeded US$2 billion, and the investment lineup behind it is not inferior to that of the world's top companies.

From the top venture capital companies in the United States such as Lightspeed Venture Capital, Redpoint Venture Capital, and Index Venture Capital, to Silicon Valley venture capital giants a16z, NVIDIA, Salesforce, BNP Paribas, etc., they all favor Mistral AI. After several rounds of financing, Mistral AI has already entered the ranks of AI unicorns.

Microsoft also announced a new partnership with Mistral AI yesterday, promising that cooperation with Mistral AI will focus on the following three key areas:

Supercomputing infrastructure: Microsoft will support Mistral AI through Azure AI supercomputing infrastructure for AI training and inference workloads.

Expanded Marketplace: Microsoft and Mistral AI will make Mistral AI’s advanced models available to customers through the MaaS and Azure Machine Learning Model Catalog in Azure AI Studio.

AI R&D: Microsoft and Mistral AI will explore collaboration to develop proprietary models for select customers, even for European public sector workloads.

But Microsoft's move was questioned by netizens. X user @osxzxso insinuated that Microsoft is trying to adopt an "intricate" strategy to monopolize the market. Musk also expressed his agreement in the comment section with a dumbfounding emoji.

Spreading Nutella on both sides of the bread I see
I saw peanut butter spread on both sides of the bread

It is worth mentioning that careful netizens have observed that after Mistral AI launched the new model, the relevant content on its official website about its commitment to the open source community has quietly disappeared, and the new model does not support open source.

However, Mistral CEO Mensch said in an interview with the Wall Street Journal that Mistral AI has not violated the original intention of open source, but has adopted a dual strategy of simultaneously promoting commercialization strategies and maintaining open source commitments.

Clearly, we need to find a fine balance between building a business model and maintaining our open source values. We want to invent new things and new architectures, but also provide our customers with more salable products.

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo