Mistral: Mistral 7B Instruct

Mistral 7B Instruct is the instruction-tuned version of Mistral 7B, the 7.3-billion-parameter model that Mistral AI announced on September 27, 2023 under the Apache 2.0 license. Mistral published three versions of the instruct model; the v0.3 model card describes it as an instruct fine-tune of Mistral-7B-v0.3. On Mistral's hosted API the model is retired: the deprecated models table lists all three versions of Mistral 7B (open-mistral-7b) as deprecated on 11/30/2024 and retired on 3/30/2025, with Ministral 3 8B as the replacement. The open weights remain on Hugging Face. Mistral's current model list is at mistral.ai/models.

Key Features

• Efficient architecture: according to Mistral's announcement, the model uses grouped-query attention for faster inference and sliding window attention, where each layer attends to the previous 4,096 hidden states.
• Vendor comparisons: Mistral said the base model beats Llama 2 13B on all benchmarks it tested and performs like a Llama 2 model more than three times its size on MMLU. These are the vendor's claims from 2023.
• v0.3 changes: the card lists an extended vocabulary of 32,768 tokens, support for the v3 tokenizer and function calling, compared with v0.2.
• License: Apache-2.0, with no usage restrictions beyond the license text.

Use Cases

• Local or on-device chat and text generation on modest hardware.
• A base for fine-tuning and function-calling experiments with an open model.
• Understanding the lineage of later Mistral small models.

Limitation: it is a 2023-era model, its card states no context length, and the hosted API version is retired, so you must self-host the weights.

Pricing and availability

The weights are free under Apache 2.0. The hosted open-mistral-7b endpoint was retired on March 30, 2025, so there is no current first-party API price. For hosted options see Mistral Small.

Getting Started

1. Download mistralai/Mistral-7B-Instruct-v0.3 from Hugging Face.
2. Use Transformers 4.42.0 or later for function calling. Tool call ids must be exactly 9 alphanumeric characters.
3. For hosted use, pick a current model from Mistral's list instead.

FAQ

Is Mistral 7B Instruct still available?

Weights are on Hugging Face under Apache 2.0, but Mistral's own API retired the 7B models on 3/30/2025.

What replaced Mistral 7B on the API?

Mistral's table lists Ministral 3 8B as the replacement for all three versions.

Is Mistral 7B Instruct open source?

Yes, it is licensed under Apache 2.0.

What was the older Mistral Tiny?

Mistral Tiny was the API name for the same 7B model; see Mistral Tiny.

Alternatives

• Mistral Tiny: the old API alias for the 7B model.
• Mistral Small: Mistral's larger small model.
• Mistral Nemo: a 12B model built with NVIDIA.

Conclusion

Mistral 7B Instruct was an influential small open model and is still usable as downloaded weights, but the API model is retired. See more in the models category.