Mistral 7B Model Card

Mistral-7B-Instruct-v0.3 is an instruct fine-tuned large language model released by Mistral AI, based on the Mistral-7B-v0.3 base model. It is the third instruct revision in the Mistral 7B series, superseding Mistral-7B-Instruct-v0.2.

The model supports text generation, instruction following, and function calling via a structured tool-call interface requiring 9-character alphanumeric tool call IDs. It uses the v3 tokenizer with a vocabulary extended to 32,768 tokens. No benchmark scores or context window size are disclosed in this document.

Not disclosed in this document.

Not disclosed in this document. The card does not describe any red-teaming, adversarial evaluation, or catastrophic-risk assessments.

The card states explicitly that the model "does not have any moderation mechanisms." No classifiers, refusal training, content filters, or access-tier controls are described.

The model is released under the Apache 2.0 license and is available on Hugging Face. It is compatible with the mistral-inference library and Hugging Face Transformers (version 4.42.0 or higher required for function calling). Inference is marked as disabled in the model card metadata, indicating hosted inference is not provided directly via Hugging Face.

The card describes the model as "a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance" and states it has no moderation mechanisms. Mistral AI notes it is "looking forward to engaging with the community on ways to make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs."

Compared to v0.2, v0.3 introduces three changes: vocabulary extended to 32,768 tokens, support for the v3 tokenizer, and support for function calling.