avatarSalvatore Raieli

Summary

A new open-source model named MiQu has emerged on HuggingFace, sparking excitement and speculation within the AI community due to its performance that rivals GPT-4, with connections to the successful AI start-up Mistral.

Abstract

The AI community is abuzz with the sudden appearance of MiQu, a 70B model on HuggingFace, which has shown performance comparable to GPT-4 on EQ-Bench. This mysterious model has prompted discussions and comparisons with Mistral, a prominent European AI start-up known for its open-source contributions and significant funding. Mistral's previous models, such as the Mistral 7B and 8x7B, have been well-received, and the community is intrigued by the possibility that MiQu could be a quantized version of a Mistral model. The model's unexpected release, potentially by an employee of Mistral or an early access customer, has raised questions about the intent behind its availability on HuggingFace. The situation is particularly noteworthy as it underscores the potential of open-source models to compete with proprietary giants like OpenAI and Google, offering state-of-the-art capabilities for free and challenging the status quo of AI development.

Opinions

  • The AI community is excited and intrigued by the performance of the MiQu model, considering it a potential rival to GPT-4.
  • Some speculate that MiQu could be a quantized version of a Mistral model, possibly released by an over-enthusiastic employee or in an unconventional manner, as seen with previous Mistral releases.
  • The release of MiQu has significant implications for the AI industry, demonstrating that open-source models can achieve state-of-the-art results and could potentially disrupt the dominance of large proprietary models.
  • The incident highlights the importance of open-source contributions in democratizing access to cutting-edge AI technologies.
  • The community is divided on whether the release was intentional or accidental, but the presence of the model on HuggingFace, despite the lack of immediate action to remove it, suggests a possible tacit endorsement of its availability.

| ARTIFICIAL INTELLIGENCE| LLMs | AI|

MiQu: Can a mysterious model be a GPT-4 rival?

An open-source model seems to be performing as GPT-4 but we do not know much about it

Photo by Johannes Plenio on Unsplash

A new model appears on HuggingFace and immediately something strange is noticed. An unsolved mystery that inflames the data scientist community until official communication arrives. What is MiQu? Who built it? And why so much excitement?

In short, we discuss it here.

The mysterious model

screenshot by the author. image source: here

The open-source community has been shaken by a small earthquake, and this one is the epicenter. Who posted this model? More importantly, what model would it be?

A 70B model appeared on HuggingFace and immediately caught the eye. Why? Some users noticed that the prompt format is the same as Mistral.

image source: here

To recap. In recent months Mistral has emerged as the leading open-source alternative to OpenAI. For the few who don’t know, Mistral is a Paris-based start-up founded by a few researchers who came out of large companies. Mistral in a few months became a unicorn and received significant funding. Mistral recently became one of the richest European start-ups in December with nearly $500 million raised (valued at nearly $2B to date).

Mistral AI continues its mission to deliver the best open models to the developer community. Moving forward in AI requires taking new technological turns beyond reusing well-known architectures and training paradigms. Most importantly, it requires making the community benefit from original models to foster new inventions and usages. (source)

From what we read on their website, Mistral is intent on revolutionizing AI by releasing models in open source. So far, it has released two models that have been hugely successful:

Returning to Miqu, it became apparent that something was out of the norm because a link was posted on 4Chan. Those who remember, when LLaMA was published one could only access the model weights after filling out a form on the META site. Until someone released the weights on 4Chan.

Ok, weird, but why so much interest?

Because it appears that the model is not simply an LLM, but something capable of rivaling GPT-4 on EQ-Bench. Clearly, this is a bomb.

screenshot by the author. source

The mystery deepens because this model has similar results to a model called Mistral Medium found on Perplexity. According to some, this model might stand for MIstral QUantized (or MiQu for friends).

So according to some, this model is a Mistral model that has been quantized. Quantization is a machine learning technique where you reduce the weight of a model at the cost of losing some accuracy.

The question remains whether it was Mistral who secretly released it or someone in the company who decided to release it without consent.

In the end, it seems the CEO of HuggingFace confirmed that it was an employee:

An over-enthusiastic employee of one of our early access customers leaked a quantised (and watermarked) version of an old model we trained and distributed quite openly.

He also seems to have taken it well, since he also commented like this

Parting Thoughts

Photo by Saif71.com on Unsplash

We don’t know if it was intentional or not, though, the model is still on HuggingFace and the company does not seem intent on removing it. On the other hand, Mistral is no stranger to releasing models in unusual ways; the previous model Mixtral 8x7B was released via torrent.

Whether intended or not, this model looks extremely promising. We don’t know if an unquantized or fine-tuned version will be able to beat GPT-4, though if it did it would be a record for open-source (and certainly problematic for companies like Google or OpenAI). Why?

Well, because a model that reaches the state of the art would be available to everyone and for free. This also has a more political significance: the open-source community can get models that can play catch-up with the best proprietary models.

What do you think about it? Let me know in the comments!

If you have found this interesting:

You can look for my other articles, and you can also connect or reach me on LinkedIn. Check this repository containing weekly updated ML & AI news. I am open to collaborations and projects and you can reach me on LinkedIn.

Here is the link to my GitHub repository, where I am collecting code and many resources related to machine learning, artificial intelligence, and more.

or you may be interested in one of my recent articles:

Artificial Intelligence
Machine Learning
Technology
Data Science
Programming
Recommended from ReadMedium
avatarChristopher Tao
Do Not Use LLM or Generative AI For These Use Cases

Choose correct AI techniques for the right use case families

7 min read
avatarVipra Singh
LLM Architectures Explained: NLP Fundamentals (Part 1)

Deep Dive into the architecture & building of real-world applications leveraging NLP Models starting from RNN to the Transformers.

39 min read
avatarThuwarakesh Murallie
How to Build Helpful RAGs with Query Routing.

An LLM can handle general routing. Semantic search can handle private data better. Which one should you pick?

10 min read
avatarFabio Matricardi
This AI breaks all records

Gemma2–2B is the raising star, outsmarting 175 Billion competitor. Here is how you can run it Locally on your PC

14 min read
avatarKaran Singh
Calculate : How much GPU Memory you need to serve any LLM ?

Just tell me how much GPU Memory do i need to serve my LLM ? Anyone else looking for this answer ? Read On …

3 min read
avatarJim the AI Whisperer
How employers are setting traps to spot AI-generated job applications and trip them up

The clever prompt HR managers can hide in job postings to catch AI

8 min read