Mixtral 8x22B

Model Overview

Mixtral 8x22B is Mistral AI’s open-weights sparse mixture-of-experts (MoE) model, released on April 10, 2024. The architecture combines 8 expert networks of 22B parameters each, totaling 141B parameters but activating only ~39B per token. This design delivers performance comparable to dense 70B+ models at roughly one-third the inference cost. The model supports a 64K context window and handles English, French, Italian, German, and Spanish with strong multilingual capability. Released under Apache 2.0, Mixtral 8x22B is one of the few frontier-class models available for unrestricted commercial use, including fine-tuning, redistribution, and integration into proprietary products. It is widely deployed through inference providers such as Together AI, Fireworks AI, and self-hosted on consumer-accessible GPU clusters.

Key Specifications

Benchmark Performance

Pricing

Pricing source: https://mistral.ai/technology/ (as of 2024-08-01). Apache 2.0 license allows unrestricted self-hosting.

Strengths

• Apache 2.0 license enables unrestricted commercial use, including redistribution.
• MoE architecture delivers strong performance-per-FLOP at inference time.
• Strong multilingual capability across five major European languages.
• Lower inference cost than dense models of comparable size.

Weaknesses

• Coding performance (HumanEval 45.2) significantly trails flagship models.
• 64K context window is smaller than newer open-weights competitors.
• MoE serving requires careful routing and memory management for optimal throughput.

Use Cases

• Cost-sensitive multilingual text generation in European markets.
• On-premise customer service or RAG pipelines with permissive licensing.
• Research and education use cases requiring Apache 2.0 compliance.
• Foundation for fine-tuned domain-specific variants.

References

• Mixtral 8x22B Announcement — Mistral AI
• Mistral AI Documentation
• Mistral Technology Page