Activefrontierllm Open Source

Mixtral 8x22B

by Mistral AI· Released April 2024· Cutoff March 2024

Mixtral 8x22B is a sparse mixture-of-experts (MoE) model with 141 billion total parameters, of which 39 billion are active per token. It is Mistral AI's largest open-weight model, offering strong performance across reasoning, coding, and multilingual tasks while maintaining efficiency through its MoE architecture.

Official Site API Docs 🤗 Hugging Face

Input cost

Free (open source)

Output cost

Free (open source)

Context window

65K tokens

Max output

—

Modalities

text

Parameters

141B (39B active)

License

Apache-2.0

Capabilities

Function CallingCode GenerationMultilingual SupportReasoningJSON ModeStreaming

Best For

Complex reasoning, coding, and multilingual tasks requiring high performance with efficient inference.

Strengths

Strong reasoning and coding capabilities, competitive with larger dense models
Efficient inference due to MoE architecture (only 39B active parameters)
Multilingual proficiency across English, French, German, Spanish, Italian
Open-weight release under Apache 2.0 license

Limitations

Smaller context window (65K) compared to some competitors (128K+)
No native vision or multimodal support
Higher memory requirements than smaller models due to total parameter count
Not as widely optimized in third-party inference stacks as some dense models

Use Cases

Advanced code generation and debugging

Complex mathematical and logical reasoning

Multilingual content creation and translation

Data analysis and summarization

Chatbots and virtual assistants

Educational tutoring and explanation

Research and scientific literature analysis

Improvements Over Previous Model

Increased total parameters from 46.7B (Mixtral 8x7B) to 141B, with active parameters rising from 12.9B to 39B
Context window expanded from 32K to 65K tokens
Significantly improved performance on benchmarks: MMLU 79.4% vs 70.6%, HumanEval 75.6% vs 63.2%
Enhanced multilingual capabilities with stronger performance on non-English tasks
Larger capacity for complex reasoning and code generation tasks

Back to all models

Improvements Over Previous Model

Increased total parameters from 46.7B (Mixtral 8x7B) to 141B, with active parameters rising from 12.9B to 39B

Context window expanded from 32K to 65K tokens

Significantly improved performance on benchmarks: MMLU 79.4% vs 70.6%, HumanEval 75.6% vs 63.2%

Enhanced multilingual capabilities with stronger performance on non-English tasks

Larger capacity for complex reasoning and code generation tasks