Mixtral 8x22B
Sparse mixture-of-experts open model with 39B active parameters out of 141B
Mixtral 8x22B is an open-weight large language model built by Mistral AI as a sparse Mixture-of-Experts, using 39B active parameters out of 141B total. Released under Apache 2.0, it targets reasoning, maths, coding and multilingual work with a 64K token context window.
- 141B total parameters, only 39B active per token
- Released under the Apache 2.0 licence
- 64K token context window
- Instructed version: 90.8% on GSM8K maj@8
- Fluent in English, French, Italian, German and Spanish
What it is
Mixtral 8x22B is a pretrained generative large language model from Mistral AI, built as a sparse Mixture-of-Experts. That design means only part of the network works on each token: 39B parameters are active out of 141B in total. Mistral AI positions it as an open model for reasoning, mathematics, coding and multilingual tasks, with a 64K token context window for pulling detail out of long documents. A base version is published so developers can fine-tune it, and an instructed version was released alongside it.
How it came about
Mistral AI presents Mixtral 8x22B as the next step in its open model family, following Mistral 7B and Mixtral 8x7B. The weights first appeared through a torrent release, and a repository on Hugging Face carries a version compatible with vLLM serving and the transformers library, with a different file format and parameter names from the torrent files. The company says it believes openness and wide distribution help innovation and collaboration in AI.
Inside the model
The model is a sparse Mixture-of-Experts: a network split into expert sub-networks, of which only a subset is used for each token. Here that gives 39B active parameters out of 141B, which Mistral AI says makes it faster than dense 70B models while costing less to run. It handles English, French, Italian, German and Spanish, supports function calling natively, and reads up to 64K tokens of context. The base model ships without any moderation mechanisms, since it is a pretrained checkpoint meant for further fine-tuning.
What it can do
Mistral AI reports strong results on reasoning, knowledge, coding and maths benchmarks, including MMLU, HellaSwag, WinoGrande, ARC Challenge, TriviaQA, NaturalQuestions, HumanEval, MBPP and GSM8K. The instructed version scores 90.8% on GSM8K maj@8 and 44.6% on MATH maj@4. On HellaSwag, ARC Challenge and MMLU in French, German, Spanish and Italian, the company says it clearly beats LLaMA 2 70B. Figures are Mistral AI's own comparisons against other open models.
How to use it
Weights are published under Apache 2.0, which Mistral AI describes as the most permissive open licence, allowing use anywhere without restrictions. The Hugging Face repository works with vLLM and the transformers library, and the model is also served on la Plateforme with a constrained output mode for function calling. The base model is offered as a starting point for fine-tuning.
Safety and testing
The published base model carries no moderation mechanisms of its own, as it is a raw pretrained checkpoint. No further safety evaluation or red-teaming detail is given here.
Why it matters
Mistral AI places Mixtral 8x22B in a line with Mistral 7B and Mixtral 8x7B, arguing that sparse activation gives a better performance-to-cost ratio than dense models of similar size. The company frames the Apache 2.0 release as a way to widen access and support application development and modernisation of technology stacks.
Questions people ask about Mixtral 8x22B
What does sparse Mixture-of-Experts mean here?
The network is divided into expert sub-networks and only some of them run for each token. In Mixtral 8x22B that means 39B active parameters out of 141B in total, which Mistral AI says lowers the cost of running the model compared with a dense one of similar size.
How can I use it?
Weights are downloadable under Apache 2.0, which Mistral AI describes as allowing use anywhere without restrictions. The Hugging Face build works with vLLM and the transformers library, and the model is also available through la Plateforme with a constrained output mode for function calling.
What is it good at?
Mistral AI highlights reasoning, mathematics, coding and multilingual use across English, French, Italian, German and Spanish. The instructed version reaches 90.8% on GSM8K maj@8 and 44.6% on MATH maj@4 in the company's own reporting.
Does it have safety filters built in?
No. The published checkpoint is a pretrained base model and, as stated on its model card, it carries no moderation mechanisms. It is intended as a basis for further fine-tuning.
Good reading and sources
There is no English Wikipedia article on Mixtral 8x22B yet, so the facts on this page come from the sources listed above: mistral.ai/news/mixtral-8x22b, huggingface.co/mistralai/Mixtral-8x22B-v0.1. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.