Sponsored by Byond Boundrys Consulting - Empowering Ideas, Delivering Results
Mistral AI Advanced

Mixtral 8x22B

Mixtral 8x22B is Mistral's Apache 2.0 MoE model with 64K context. Its managed API was retired on March 30, 2025, though weights remain available.

Language Models (LLMs)Text Paid
In plain English

What is this model and why does it matter?

Mixtral 8x22B is Mistral's 141B-parameter MoE family with 39B active parameters and 64K context. The API model was retired in March 2025.

Historical Mistral researchLarge open-source self-hostingMoE studyExisting compatibility deployments
Model overview

Mixtral 8x22B: features, use cases and important details

Mixtral 8x22B is Mistral AI’s large mixture-of-experts model released on April 17, 2024. Mistral’s official model page lists 141 billion total parameters, 39 billion active parameters, a 64K context window and Apache 2.0 weights for base and instruction-tuned variants.

What made Mixtral 8x22B notable

The MoE architecture activates only part of the total parameter set for each token, allowing the model to combine large overall capacity with more efficient computation than a similarly sized dense model. Mistral also documented function-calling support for the instruction workflow.

Retired API status

Mistral now marks Mixtral 8x22B as retired. The official retirement date is March 30, 2025, and Mistral recommends Mistral Small 4 for new integrations. The old API identifier open-mixtral-8x22b therefore should not be shown as an active current API, and historical token prices should be cleared.

Why keep this page

The model remains useful as an open historical checkpoint because its Apache 2.0 weights can still be studied and self-hosted. It is especially relevant for MoE research, reproducibility and existing deployments that depend on the older model. New managed deployments should follow Mistral’s current model guidance.

Get started

How to use this model

  1. Use the official Mistral model page to choose base or instruct weights.
  2. Download the Apache 2.0 checkpoint if self-hosting.
  3. Provision substantial GPU memory for the 141B-parameter MoE model.
  4. Use the correct tokenizer and chat template for the instruct variant.
  5. For new managed API integrations, follow Mistral's recommended replacement instead.
Copy and try

Example prompts

  • Analyze this multilingual policy document and summarize the tradeoffs.
  • Use the available tools to retrieve data and explain the result.
  • Review this architecture proposal and identify scaling risks.
Capabilities

What it can do

  • Mixture-of-experts text generation
  • 64K context
  • Multilingual text
  • Instruction following
  • Function calling in the instruct/API workflow
Best for

Practical use cases

  • MoE research
  • Large self-hosted assistants
  • Historical Mistral compatibility
  • Multilingual analysis
Pricing

What does it cost?

Retired Mistral API model with Apache 2.0 weights still available. Historical API prices should not be treated as current.

Simple summaryApache 2.0 weights remain available for self-hosting, but the retired Mistral API should not be assigned current token pricing.

What stands out

  • Apache 2.0 license
  • 64K context
  • Only 39B active parameters per token despite 141B total
  • Official base and instruct weights

Things to consider

  • Retired API model
  • Very large hardware footprint
  • Newer Mistral models are recommended
Limitations

Important restrictions and trade-offs

  • Not appropriate as a new Mistral API integration
  • Self-hosting is resource intensive
  • Historical benchmark performance should not be confused with current models
SimplifyAITools verdict

Our editorial take

Keep as a valuable historical open model page, but make the retirement unmistakable. Mistral explicitly recommends a newer model for new integrations.

References

Primary sources

  1. Open source 1 ↗
  2. Open source 2 ↗
  3. Open source 3 ↗
Keep exploring

Related AI models