Mixtral 8x22B Review: Mistral’s Largest Open-Weight MoE Model (April 2024)

Published April 17, 2024 — Updated June 15, 2026

On April 17, 2024, Mistral AI released Mixtral 8x22B — their largest open-weight model to date and the first sparse mixture-of-experts (MoE) model to seriously challenge closed frontier models on reasoning and coding benchmarks. It was released under the Apache 2.0 license, making it free for commercial use and self-hosting.

What Is Mixtral 8x22B?

Mixtral 8x22B is a sparse MoE model. Unlike dense language models that activate all parameters for every token, it routes each token through only 2 of 8 expert sub-networks (each 22B parameters). The result is 141B total parameters but only ~39B active per token — giving it efficiency closer to a 39B dense model while retaining the capacity of a much larger one.

This architecture follows Mistral’s earlier Mixtral 8x7B (which used 7B expert groups) but scales to a level that, at launch, outperformed every other publicly available open-weight model on standard benchmarks.

Key Specs

• Architecture: Sparse MoE — 8 expert groups of 22B parameters, 2 active per token
• Total parameters: ~141B
• Active parameters per token: ~39B
• Context window: 64,536 tokens
• Languages: English, French, German, Italian, Spanish
• Capabilities: Text generation, function calling, JSON mode, multilingual
• License: Apache 2.0
• Weights: Hugging Face — mistralai/Mixtral-8x22B-Instruct-v0.1

Benchmarks vs LLaMA 3 70B and GPT-3.5

At launch, Mixtral 8x22B was the strongest open-weight model on coding and math, while trading blows with LLaMA 3 70B on general knowledge.

What’s New vs Mixtral 8x7B?

• 3× larger experts: Expert groups grow from 7B to 22B parameters, significantly increasing capacity
• 8× larger context: 64k tokens vs 8k on Mixtral 8x7B — enough to fit entire codebases or legal documents
• Function calling: Native support added, enabling agent and tool-use workflows
• JSON mode: Reliable structured output without prompt hacks
• Math uplift: MATH benchmark jumps from ~28% (8x7B) to 41.8%

Pricing

Mixtral 8x22B is available two ways:

• Self-hosted (free): Download the Apache 2.0 weights from Hugging Face. Requires ~4× A100 80GB GPUs or equivalent. Works with vLLM, TGI, llama.cpp, and Ollama.
• Mistral API: Available via La Plateforme as open-mixtral-8x22b — approximately $2/M input and $6/M output tokens at launch. Mistral has since released newer, cheaper models that handle most workloads better.

Who Should Use It?

• Teams running self-hosted inference that want the best open-weight model from the 2024 vintage without licensing restrictions
• Researchers benchmarking or fine-tuning large MoE architectures — the Apache 2.0 license makes this straightforward
• Multilingual applications across English, French, German, Italian, and Spanish
• Long-document pipelines that need a 64k context window without paying frontier model prices

Is Mixtral 8x22B Still Relevant in 2026?

For most new projects, Mistral Small 4 or Codestral 25.08 are better choices — they are newer, more capable, and still Apache 2.0. However, Mixtral 8x22B remains useful as a well-understood research baseline, is already deployed in many production systems, and benefits from broad community support across inference frameworks.

If you are already running Mixtral 8x22B in production, there is rarely an urgent reason to migrate unless you need vision, a 256k context window, or the additional benchmark improvements in newer models.

Frequently Asked Questions