Mixtral 8x22B: Complete Specifications, Pricing, API Access & Use Cases (2026)
What Is Mixtral 8x22B?
Mixtral 8x22B is Mistral AI’s open sparse Mixture-of-Experts large language model, released on April 17, 2024, featuring a 64K-token context window and text-generation capabilities, with current hosted API pricing not listed because Mistral documentation marks the model retired as of June 2026.
Mistral described Mixtral 8x22B as a sparse MoE model with 141B total parameters and 39B active parameters per token, designed to offer a larger open-weight alternative to dense general-purpose LLMs. The model is relevant for users researching open-weight multilingual generation, coding, math reasoning, long-document prompting, function calling, and self-hosted LLM deployment.
The model’s current production status needs careful interpretation. It was available through Mistral’s hosted platform at launch, but Mistral’s current model card lists a March 30, 2025 retirement date and points to Mistral Small 3.2 as the replacement. Teams comparing legacy open Mistral models with newer hosted options often evaluate Mistral Large API access or newer Mistral Small 4 specifications instead.
What Are Mixtral 8x22B’s Key Specifications and Pricing?
| Field | Verified Value |
|---|---|
| Provider | Mistral AI (as of June 2026) |
| Model Family | Mixtral / Mistral open models (as of June 2026) |
| Model Type | Open sparse Mixture-of-Experts large language model (as of June 2026) |
| Release Date | April 17, 2024 (as of June 2026) |
| Context Window | 64K tokens (as of June 2026) |
| Input Pricing | Not listed for current hosted API use because the Mistral-hosted model is retired (as of June 2026) |
| Cached Input Pricing | Not confirmed from official sources for current Mixtral 8x22B hosted use as of June 2026 |
| Output Pricing | Not listed for current hosted API use because the Mistral-hosted model is retired (as of June 2026) |
| Pricing Unit | Mistral prices current API models per million tokens, but Mixtral 8x22B is not listed as a current priced API model (as of June 2026) |
| Modality Support | Text-focused LLM; image, audio, and video input are not confirmed for Mixtral 8x22B in official documentation as of June 2026 |
| Supported Input Types | Text prompts/messages (as of June 2026) |
| Supported Output Types | Text output (as of June 2026) |
| API Access | Historical Mistral hosted access; current Mistral model card lists retirement on March 30, 2025 (as of June 2026) |
| Model ID | open-mixtral-8x22b; Hugging Face instruct ID: mistralai/Mixtral-8x22B-Instruct-v0.1 (as of June 2026) |
| Availability | Open weights remain documented; hosted Mistral API model is retired (as of June 2026) |
| Knowledge Cutoff | Not confirmed from official sources as of June 2026 |
| Rate Limits | Not confirmed for Mixtral 8x22B current hosted use as of June 2026 |
| Fine-tuning Support | Weights may be used under Apache 2.0 terms; current hosted fine-tuning support is not confirmed for Mixtral 8x22B as of June 2026 |
| Streaming Support | Not applicable to current hosted Mixtral 8x22B API status; Mistral’s general chat endpoint supports streaming for available models (as of June 2026) |
| Batch API Support | Not confirmed for current Mixtral 8x22B hosted use as of June 2026 |
| Tool / Function Calling | Native function calling was stated in Mistral’s launch material; Hugging Face documents function-calling-related tokens and examples (as of June 2026) |
| Structured Output / JSON Mode | Launch material referenced constrained output mode on la Plateforme; current Mixtral 8x22B hosted access is retired (as of June 2026) |
| License / Usage Restrictions | Apache 2.0 weights (as of June 2026) |
What Can Mixtral 8x22B Do That Makes It Useful in Production?
- Long-context text work. Mixtral 8x22B’s 64K-token context window made it suitable for large-document summarization, retrieval-augmented drafting, and multi-file analysis when teams could self-host the model or use a supported deployment. Context length still has practical limits: longer prompts increase memory, latency, and evaluation complexity.
- Multilingual generation. Mistral highlighted English, French, Italian, German, and Spanish fluency at launch. This makes the model relevant for multilingual documentation, support drafting, and internal knowledge workflows, but organizations should still test domain-specific terminology before deployment.
- Coding and math assistance. Mistral positioned Mixtral 8x22B as strong in mathematics and coding, making it useful for code explanation, test generation, data transformation, and technical Q&A. It should not be used as an unchecked authority for production code, security-sensitive fixes, or regulated decisions.
- Function-calling workflows. The model’s native function-calling support can support agent-style workflows where a model selects structured tool calls. Self-hosted implementations need careful prompt formatting, schema validation, and logging.
What Are Mixtral 8x22B’s Supported Modalities?
| Modality | Supported? | Notes |
|---|---|---|
| Text input | Yes | Supported through prompt/message-style LLM usage. |
| Text output | Yes | The instruct model is a text-generating LLM. |
| Image input | Not confirmed | Mixtral 8x22B is not documented as a vision model. |
| Audio input | Not confirmed | No official Mixtral 8x22B audio-input support found. |
| Video input | Not confirmed | No official Mixtral 8x22B video-input support found. |
| Tool/function calls | Yes, with implementation caveats | Official launch material and Hugging Face notes reference function-calling support. |
Where Does Mixtral 8x22B Fall Short?
- The main limitation is availability: Mistral’s current model card lists Mixtral 8x22B as retired from hosted API service, so teams should not plan new provider-hosted production systems around open-mixtral-8x22b unless they have a separate verified deployment path.
- Current official provider pricing is also not available for the retired hosted API model. Any historical or third-party price should be treated separately from Mistral’s current official pricing page.
- The model is text-focused. It is not a substitute for modern multimodal systems such as newer Mistral, Gemini, GPT-4o, or Claude models that support image, audio, or richer tool ecosystems. For broad multimodal comparisons, teams may separately review Gemini 1.5 Pro specifications or GPT-4o mini access patterns.
- This is a general AI limitation and is not model-specific unless stated by Mistral: Mixtral 8x22B can hallucinate, produce unsafe or biased outputs, misunderstand long prompts, and require expert review for legal, medical, financial, security, or other high-stakes contexts.
What Is Mixtral 8x22B Best Used For?
| Use Case | Why Mixtral 8x22B May Fit | Important Limitation |
|---|---|---|
| Self-hosted open-weight LLM experimentation | Apache 2.0 weights and MoE architecture make it useful for infrastructure testing. | Requires significant GPU memory and deployment expertise. |
| Long-document text analysis | 64K context supports larger prompts than earlier Mixtral 8x7B. | Long-context reliability must be evaluated task by task. |
| Multilingual internal drafting | Launch material highlights five European languages. | Domain-specific terminology still needs testing. |
| Coding and math assistance | Mistral highlighted coding and math strengths. | Generated code and calculations require validation. |
| Tool-calling research | Model materials reference function-calling support. | Production tool use needs strict schema validation and safety controls. |
How Does Mixtral 8x22B Compare to Mistral Small 4 and Llama 3.1 70B?
| Comparison Area | Mixtral 8x22B | Mistral Small 4 | Llama 3.1 70B Instruct | Scenario Fit |
|---|---|---|---|---|
| Provider | Mistral AI | Mistral AI | Meta | Useful for comparing Mistral open MoE models with Meta open-weight text models. |
| Release status | Released April 2024; Mistral-hosted API retired March 2025 | Released March 2026; listed with current pricing | Released July 2024 | New deployments usually need current availability checks. |
| Context window | 64K | 256K | 128K | Small 4 or Llama 3.1 may fit longer-context workflows better. |
| Model design | Sparse MoE, 141B total / 39B active | Hybrid model, 119B parameters / 6.5B active | Dense 70B-class transformer | Mixtral is useful for MoE self-hosting research; Llama 3.1 is widely used for open-weight text deployments. |
| Modalities | Text-focused | Documented as multimodal in current Mistral materials | Text in / text and code out | Modern multimodal tasks favor newer models. |
| Current provider pricing | Not listed because hosted model is retired | \$0.15/M input and \$0.60/M output tokens | Pricing depends on hosting provider, not Meta’s model card | Cost comparisons depend on deployment path. |
| Access | Open weights; retired Mistral hosted model ID | Current Mistral API/model card | Open weights via Meta/Hugging Face, subject to license | Use Mixtral 8x22B when open-weight legacy MoE behavior is specifically needed. |
How Do I Access Mixtral 8x22B?
Mistral’s current documentation lists the historical provider model ID as open-mixtral-8x22b, but also lists a March 30, 2025 retirement date for hosted access.
Developers who still need Mixtral 8x22B should evaluate the open weights, the Apache 2.0 license terms, required GPU memory, tokenizer compatibility, and any third-party serving stack separately. Hugging Face lists the instruct model as mistralai/Mixtral-8x22B-Instruct-v0.1, with Apache 2.0 licensing and 141B parameters.
FAQs
What is Mixtral 8x22B’s context window?
Mixtral 8x22B has a 64K-token context window in Mistral’s official model card as of June 2026.
How much does Mixtral 8x22B cost?
Current official Mistral hosted API pricing is not listed because Mistral’s model card marks Mixtral 8x22B retired. Treat any third-party or historical price separately.
Can I use Mixtral 8x22B through the Mistral API?
Mistral’s current documentation lists the model ID open-mixtral-8x22b but marks hosted access retired on March 30, 2025.
What is Mixtral 8x22B suitable for?
It may fit self-hosted open-weight LLM experiments, multilingual text generation, long-document analysis, coding assistance, and function-calling research, provided outputs are validated.
