Mixtral 8x22B Model Card
What this is
Mixtral-8x22B-Instruct-v0.1 is an instruct fine-tuned version of Mixtral-8x22B-v0.1, released by Mistral AI with a card dated 2026-03-31. It is a sparse mixture-of-experts large language model released under the Apache-2.0 license. The card is structured primarily as an integration guide rather than a traditional model card.
Capabilities
The model supports five languages: English, Spanish, Italian, German, and French. It includes native function-calling support with tool call IDs of exactly 9 alphanumeric characters, exposed through both the mistral-common and HuggingFace transformers APIs. No benchmark scores, parameter counts, or context window sizes are stated in this document.
Evaluation methodology
Not disclosed in this document.
Safety testing
Not disclosed in this document.
Mitigations
The card references a privacy policy governing personal data processing but does not describe any content classifiers, refusal training, or safety mitigations applied to the model. No ASL or FSF tier is invoked.
Deployment and access
The model is available on HuggingFace under the Apache-2.0 license with an extra-gated repository requiring acknowledgment of Mistral AI's privacy policy. It is compatible with vLLM, mistral-inference, and HuggingFace transformers (version 4.42.0 or higher required for function calling). No API-tier restrictions, usage limits, or access-control tiers beyond the gated download are described.
Limitations
The card notes a tokenizer discrepancy: the HuggingFace tokenizer may not produce results identical to the mistral-common reference implementation, and community PRs to correct it are solicited. No other limitations are disclosed in this document.
What's new
This is the v0.1 release of the instruct variant; no prior instruct version is referenced and no changelog entries appear in the document.