After a first Mixtral 8x7B model released in torrent in December, the French startup Mistral AI is releasing another, presented as more efficient: Mixtral 8x22B. You must again use a BitTorrent client to retrieve it.
Without warning, the French startup Mistral AI has just published this Wednesday, April 10, a brand new language model — Mixtral 8x22B. The company, however, did not provide any comments to accompany this release, neither on its X account (formerly Twitter) nor on its official website – the last news dating back to February 26, 2024.
Here again, Mistral has opted for distribution via the BitTorrent protocol. The tweet about Mixtral 8x22B indeed contains a magnet link (for those who are not familiar with P2P exchanges, this will mostly seem like gibberish) allowing you to retrieve the model. You need a BitTorrent client to download it.
A 262 GB model to download
A word of caution is in order: allow plenty of space. The whole thing (in total, there are four files associated with the torrent link) weighs 261.92 gigabytes. If you plan to recover it, you will need to opt for a storage medium large enough to accommodate it – the most recent hard drives, however, reach a few terabytes.
The model is in fact three times more massive than the previous one from Mistral, which was shared in December – Mixtral 8x7B, that’s its name, weighs 87 GB. At the time, the startup had also opted to make it available via a peer-to-peer exchange. A press release was also posted online to present its main specificities.
Mistral had described Mixtral as a “ sparse mix of high-quality expert models with open weights “. According to the test bench chosen by the French company, Mixtral 8x7B exceeded GPT-3.5 from OpenAI and LLaMA 2 70B from Meta (Facebook) in several categories. These two rival models date from November 2022 and July 2023.
Google and Facebook update their models
The technical characteristics and performances of this Mixtral 8x22b are not given by Mistral and the accounts of the founders of the company, as Arthur Mensch Or William Lample, have remained silent until now. The profile of this model is intriguing in any case, as we can see on Reddit or Y Combinator.
The arrival of Mixtral 8x22b in any case comes in a fairly lively context in terms of generative artificial intelligence. For example, we learned on April 9 of the arrival in May of the LLaMA 3 language model at Meta — it too will be shared according to open source codes. Moreover, the American giant is holding a conference today in Paris.
On the side of Google, which is also increasing its efforts in generative AI, there were two notable announcements: the availability of Gemini 1.5 Pro in more than 180 countries, with several advances, and the debut of CodeGemma, a version of Gemma (a language model) which is suitable for generating computer code.
Subscribe to Numerama on Google News so you don’t miss any news!