Mistral Inference Is Archived, but the Model Weights Are Still Worth Your Time

You've probably been there: you find a promising open-source inference library, only to discover the repository has been archived. That's exactly the situation with Mistral Inference, the official inference code from Mistral AI. But before you close the tab, here's the thing—the codebase is minimal, the installation is straightforward, and the model weights are still available for download. If you want to run Mistral models locally without relying on a hosted API, this is still a perfectly functional starting point.

What It Does

Mistral Inference is exactly what it sounds like: a minimal codebase for running Mistral's family of language models on your own hardware. The repository provides a small Python package that handles model loading and inference, supporting everything from the original 7B model up through Mistral Large 2 and Pixtral 12B. The code relies on xformers under the hood, which means you'll need a GPU to install it—that's a hard requirement, not a suggestion.

The package is available via PyPI as mistral-inference, or you can clone the repo and install it locally with Poetry. The README links to download tarballs for each model variant, complete with md5 checksums so you can verify your downloads. The supported models span the full Mistral lineup: the base 7B and 8x7B models, the Mixtral mixture-of-experts variants at 8x7B and 8x22B, Codestral and Mathstral for specialized tasks, the Codestral-Mamba 7B (a Mamba architecture model), Nemo base and instruct, and Mistral Large 2.

Why It's Cool

The honest appeal here is that it's official—these are the same folks who trained the models, packaging up their own inference code. That means the implementation is likely faithful to how the models were designed to run. But there's a catch, and it's a big one: the repository is archived and no longer maintained. So why bother?

• Direct model downloads. You don't need to go through Hugging Face or any intermediary. The README provides direct tar URLs for every model, which is refreshingly straightforward. You download, verify the checksum, and run.
• Minimal surface area. The codebase is intentionally small. If you want to understand how inference works for these models without wading through a massive framework, this is a clean starting point. You can read the source in an afternoon.
• Colab-ready. There's a link to a Google Colab notebook in the README, which means you can test everything in a browser before committing to local setup. That lowers the barrier to entry significantly.
• Guardrailing support. The documentation link points to Mistral's guardrailing features, which is a practical inclusion if you're deploying something that needs content safety filtering.

The trade-off is obvious: no updates, no bug fixes, no community pull requests being merged. If you hit a problem, you're on your own. But for experimentation, learning, or running a specific model version in a controlled environment, that might be fine.

How to Try It

Assuming you have a GPU available, getting started is a few commands away. First, install the package:

pip install mistral-inference

If you prefer to work from source:

cd $HOME && git clone https://github.com/mistralai/mistral-inference
cd $HOME/mistral-inference && poetry install .

Next, pick a model and download it. The README provides direct links with checksums. For example, the 7B Instruct model:

wget https://models.mistralcdn.com/mistral-7b-v0-3/mistral-7B-Instruct-v0.3.tar
md5sum mistral-7B-Instruct-v0.3.tar
# Should match: 80b71fcb6416085bcb4efad86dfb4d52

Then extract the tarball and run the inference script. The repository's tutorials folder has a Colab notebook that walks through the full workflow, which is probably the easiest way to get oriented.

You can find the full repository at github.com/mistralai/mistral-inference.

Final Thoughts

This is a time capsule, not a living project. If you need the latest Mistral model with ongoing support and active development, you should look at whatever Mistral AI launches next or use the hosted API. But if you want to run a specific model version locally, understand the implementation, or just tinker with Mistral's architecture on your own hardware, this archived repository still works. The weights are out there, the code compiles, and sometimes that's all you need.

Follow @githubprojects for more developer tools and open source projects.