Back to Mixtral
Mistral AI /

Mixtral 8x22B
Mixtral 8x22B

Open WeightsVendor API RetiredChat & ReasoningtextcodeUpdated April 17, 2024

Model Overview

Mixtral 8x22B is Mistral AI's open-weights Sparse Mixture-of-Experts (SMoE) model. It sets a new standard for performance and efficiency within the open-source community, offering unparalleled capabilities for a model of its active parameter size.

Capabilities

  • Sparse Mixture-of-Experts (SMoE): Mixtral 8x22B uses a sparse architecture where it possesses 141 billion total parameters but only activates 39 billion parameters per token. This allows it to achieve the performance of a much larger model while maintaining the inference speed and cost of a smaller one.
  • Exceptional Reasoning: The model demonstrates outstanding capabilities in mathematics, coding, and logical reasoning, consistently outperforming other open models in its weight class.
  • Multilingual Proficiency: It is highly fluent in multiple languages, including English, French, Italian, German, and Spanish, making it suitable for global applications.
  • Function Calling: It possesses native function calling capabilities, allowing it to seamlessly integrate with external APIs and tools to execute complex workflows.

Example Use Cases

  • High-Throughput Applications: Ideal for applications requiring fast response times and low latency, such as real-time chatbots or code completion tools, due to its efficient SMoE architecture.
  • Enterprise Deployments: Its open-weights nature allows businesses to host the model securely within their own infrastructure, ensuring data privacy and control.
  • Multilingual Assistants: Building conversational agents that can fluently interact with users across different European languages.

Performance & Benchmarks

Mixtral 8x22B delivers top-tier performance for an open model:

  • MMLU (General Knowledge): 77.3%
  • HumanEval (Code): 75.0%
  • GSM8K (Math): 88.6%

Open-Weights Philosophy

Mistral AI released Mixtral 8x22B under the highly permissive Apache 2.0 license, allowing for unrestricted commercial use, modification, and distribution. This aligns with their commitment to fostering an open and collaborative AI ecosystem.

About Mistral AI

Mistral AI is a European artificial intelligence company based in Paris, focused on developing highly efficient, open, and customizable generative AI models.

Key Features

Sparse MoE architecture

Feature 01

Fully open weights with commercially permissive Apache 2.0 license

Feature 02

Native function calling support

Feature 03

You might also want to compare

Verified Sources

Tags

open-sourcemoecodingapache-2.0

Model Specs

open-weights

Parameters

8x22B MoE (~39B active)

Context Window

64K

License

Apache 2.0

Deployment

self-hostable

Resources & Links

Lineage

Model Family

Part of the Mixtral family

Only release in this line currently tracked.

Compare Specs

Compare parameters, context windows, modalities, and benchmark scores of this model side-by-side with others.

Compare Model