Mistral-7B-Instruct-v0.2 is a fine-tuned version of the Mistral-7B-v0.2 language model, designed specifically for following instructions in a conversational format. It supports a 32k token context window, enabling more detailed and longer interactions compared to its predecessor. The model is trained to respond to user prompts formatted with [INST] and [/INST] tags, and it performs well in instruction-following tasks like Q&A, summarization, and explanations. It can be used via the official mistral_common tokenizer or Hugging Face’s transformers library, and supports generation on GPUs with BF16 precision. Built on a transformer architecture without sliding-window attention, the model is optimized for fast inference and chat integration. Though it lacks moderation mechanisms, it showcases the capability of Mistral-7B as a base for further fine-tuning and safety tooling. Mistral-7B-Instruct-v0.2 is licensed under Apache 2.0 and widely used across open-source projects.
Features
- 7.24B parameter transformer model fine-tuned for instructions
- Supports 32k token context length
- Utilizes RoPE with theta = 1e6 and no sliding-window attention
- Compatible with Hugging Face Transformers and native Mistral tokenizer
- Outputs structured chat responses using [INST] instruction format
- Ideal for summarization, explanation, and Q&A tasks
- Fast inference with support for BF16 precision on CUDA
- Open-source under Apache 2.0 with widespread adoption in research