Apache 2.0 sparse MoE with 39B active parameters, strong math and coding, and efficient self-hostable inference.
Modality:
Text
PDF
model ID
mistralai/mixtral-8x22b-instruct
Output Speed *
0.00
tok/s
Intelligence Index *
5.7
/ 100
Context Window *
65536
tokens
Input price
12
Anytoken
Output price
36
Anytoken
Mixtral 8x22B Instruct: Open-Weight MoE Efficiency for Coding, Math and Multilingual Workloads
Mixtral 8x22B Instruct is Mistral's instruction-tuned sparse Mixture-of-Experts model, released under Apache 2.0. It activates roughly 39B of 141B total parameters per token, giving 70B-class capability at lower inference cost. Within Mistral's open lineup it sits above Mixtral 8x7B as the larger, more capable open-weight option. Its distinguishing traits are strong math and coding performance, native function calling, and fluency across English, French, Italian, German and Spanish. It fits teams that want a self-hostable, permissively licensed model for high-volume coding, reasoning and multilingual generation.
Start building with Mixtral 8x22B Instruct via the AnyAPI.ai unified API.
Performance
Where Mixtral 8x22B Instruct Delivers: Math, Code and Efficient Inference
Mixtral 8x22B Instruct is strongest on coding and mathematics relative to other open-weight models of its era. Mistral reports it leads comparable open models on HumanEval, MBPP and GSM8K, with an instruct GSM8K result around 90% (maj@8). Because only ~39B of 141B parameters activate per token, it runs faster than a dense 70B model at similar quality. For production, that means cost-efficient throughput on code generation and quantitative reasoning when self-hosted or served through a provider, without paying full dense-model inference cost.
Benchmarks
Mixtral 8x22B Instruct on Independent Benchmarks
Independent evaluation frames this model relative to its 2024 open-weight peers rather than current frontier models. Artificial Analysis assigns Mixtral 8x22B Instruct an estimated Intelligence Index score of 10, placing it below the current median (13) for comparable open-weight non-reasoning models. It confirms the model is text-only, is not a reasoning model, and provides direct responses without extended chain-of-thought. The practical read: competitive coding and math for its size and cost, but weaker general intelligence than newer models. Treat it as an efficient workhorse, not a frontier reasoning engine.
Output Speed
*
0.00
tok/s
Intelligence Index
*
5.7
/ 100
MMLU *
Broad world knowledge and problem-solving
54
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
33
%
HLE *
Adherence to multi-step structured instructions.
4
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
15
%
Technical Specifications
What the model supports
Mixtral 8x22B Instruct is a text-in, text-out sparse Mixture-of-Experts model with a 65,536-token context window and native function calling. The Apache 2.0 license is the most production-relevant specification: teams can deploy, fine-tune and self-host anywhere without usage restrictions, which matters for data-sensitive or on-premise workloads. The 65K context supports recall across large documents, though it is smaller than newer long-context models. It is not a reasoning model and offers no extended-thinking mode; responses are direct.
Verified Specifications —
Mixtral 8x22B Instruct
*
Input modalities
Text
PDF
output modalities
Text
Context window
65536
tokens
Maximum output tokens
52428
Reasoning
No
Knowledge cutoff
April 2024
Pricing (standard)
12
AnyTokens in
/
36
AnyTokens out
Limitations & Trade-offs
1
Below-average general intelligence today. Independent evaluation (Artificial Analysis) places its estimated Intelligence Index at 10, under the current peer median of 13. The model reflects 2024-era capability, so for complex reasoning, agentic planning or nuanced instruction following, newer models will outperform it. Reserve Mixtral 8x22B Instruct for well-scoped coding, math and multilingual generation rather than frontier reasoning.
2
No reasoning mode. Mixtral 8x22B Instruct is not a reasoning model and produces direct responses without extended chain-of-thought. For multi-step problem solving, verification-heavy math, or agent trajectories that benefit from explicit deliberation, a dedicated reasoning model is preferable. You can prompt for step-by-step output, but there is no native extended-thinking configuration.
3
Text-only modality. The model accepts and produces text only; it does not support image, audio, video or document-image input. Any workload requiring vision, OCR-style document understanding from images, or multimodal grounding must use a different model. This limits it to pure text and code pipelines.
4
Context smaller than modern long-context models. At 65,536 tokens the window handles large documents but falls well short of the 128K-to-1M windows offered by newer models. For very long codebases, extended multi-document RAG or book-length inputs, a larger-context model is the better fit.
Best-Fit Workloads
Where this model earns its place
01
Code generation and completion
Mistral reports Mixtral 8x22B leads comparable open models on HumanEval and MBPP, making it a practical open-weight option for code generation, refactoring and completion. Native function calling lets it integrate directly with developer tooling. Self-hostability under Apache 2.0 suits teams that must keep source code inside their own environment.
02
Mathematical and quantitative reasoning
The instruct variant reaches roughly 90% on GSM8K (maj@8) in Mistral's reporting, indicating strong arithmetic and word-problem performance for its class. This fits calculation-heavy assistants, data-analysis helpers and educational math tooling. Note that it is not a reasoning model, so pair it with verification for high-stakes computation.
03
Multilingual generation across European languages
The model is fluent in English, French, Italian, German and Spanish, with published multilingual results on HellaSwag, Arc Challenge and MMLU across those languages. This makes it suitable for European-market content generation, localization support and multilingual chat where an open-weight, self-hostable model is preferred over a proprietary API.
04
Cost-efficient high-volume text generation
Sparse activation of ~39B of 141B parameters yields faster inference than a dense 70B model at comparable quality, giving a strong performance-to-cost ratio for its size. For high-volume generation pipelines where budget and throughput dominate, it is more economical than running an equivalently capable dense model, especially when self-hosted.
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04