Mistral Large 4: Can Australian Businesses Self-Host It?

Mistral Large 4 is now an open-weight model, which means Australian businesses can download it and run it on infrastructure they choose instead of sending every prompt to an overseas AI provider. French AI company Mistral released the weights on 29 October 2026, after a public preview earlier in October, and the model card says it is the company’s largest and most capable model to date. The catch is that “can download” is not the same as “can run in the office”. Here is what was released, what the licence allows, what the hardware really looks like and how to decide whether it belongs in your AI stack.

What Mistral released with Mistral Large 4

According to the official model card, Mistral Large 4 (nicknamed “Le Chonk”) is a mixture-of-experts model with about 1 trillion total parameters, of which around 49 billion are active for each token. It accepts text and images, writes text, supports a context window of up to 1 million tokens and is fluent in more than 160 languages. It has three reasoning levels (low, medium and high) so the same model can give a quick reply or work through a long agent task, plus native function calling and strict JSON output for software integrations.

In its release announcement, Mistral positions the model for coding agents, defensive cybersecurity and enterprise work, and says it was trained from scratch on its own infrastructure in Europe. The company reports that it effectively ties the leading open-weight model, GLM-5.3, on Artificial Analysis’s Coding Agent Index while using 47% fewer tokens per task. Hugging Face ranks it fifth among open-weight models on the Artificial Analysis Intelligence Index as at 29 October, and the model card notes that leading Chinese open-weight models score higher on that overall index. As TechCrunch reports, Mistral’s pitch is control: organisations can deploy the model themselves rather than rely on an outside AI company for access. Treat the benchmark figures as vendor-reported until independent testing catches up.

What the Mistral Large 4 licence allows

The model ships under the Mistral Ecosystem Open-Weight License, a modified MIT licence. In plain terms:

• Internal use is free. You can use, copy, modify and fine-tune the model, and internal use that does not make it available to third parties is exempt from the commercial licence requirement.
• Features inside your product are generally fine. End-user products where model capabilities are embedded within specific features are excluded from the definition of “Model as a Service”.
• Selling access to the model needs a separate licence. If your business operates a Model as a Service offering, meaning third parties get access for inference or training in a way that gives them meaningful control over inputs, parameters or training data (for example through an API or hosted endpoint), you must sign a commercial licence with Mistral before commercial use.
• Keep the notice and respect others’ rights. The copyright and permission notice must travel with copies of the model, and the model card says it must not be used in a way that infringes third-party rights.

For most businesses building internal assistants or AI features into their own software, that is a permissive starting point. If you are a software company planning to let customers send their own prompts or training data to a hosted model, have the licence reviewed before you build.

The hardware reality of self-hosting

A trillion-parameter model is not something you run on a spare desktop. The serving recipes on the model card use a node of eight GPUs, and Hugging Face’s ready-made endpoint recipe runs on eight NVIDIA H200 GPUs. The smaller NVFP4 version targets NVIDIA’s newer Blackwell GPUs, and the Dell Enterprise Hub example deploys it on four NVIDIA B300 GPUs, which is still a data centre server. Hugging Face also tested heavily compressed 2-bit versions: on a single Apple Silicon Ultra machine with 512 GB of unified memory it measured about 15 tokens per second of generation, which suits a pilot or a single user rather than a busy team. Heavier compression also trades away some quality, so test it on your own tasks.

So “self-hosted” for most Australian organisations will mean dedicated GPU servers in a data centre or rented cloud GPUs, not a box in the comms cupboard. Smaller open-weight models remain the practical choice for laptops and on-premise servers, which our free local LLM for business course covers.

Four ways to use Mistral Large 4

1. Mistral’s own API. The simplest way to test it. Mistral lists pricing of $0.70 per million input tokens and $2.10 per million output tokens, with cheaper cached input.
2. A managed cloud platform. Mistral says the model is available through Microsoft Foundry, with private deployment options for Azure developers, as well as OpenRouter and NVIDIA-hosted endpoints.
3. Rented GPU endpoints. Services such as Hugging Face Inference Endpoints let you run the open weights on dedicated hardware without buying servers.
4. Your own infrastructure. Download the weights and serve them with tools such as vLLM or SGLang on GPUs you own or lease, for full control over where data goes and which version runs.

Each option processes your data in a different place, so check the region and data handling terms rather than assuming an Australian location. Our guide to private LLM options in Australia compares these deployment patterns, and our earlier piece on sovereign AI and open-weight models explains why control is on more boards’ agendas.

When Mistral Large 4 makes sense for an Australian business

An open-weight frontier model is most useful when one of these is true:

• Data sensitivity rules out some APIs. Client files, health or financial records and source code you cannot send to a shared service are the main reason to consider self-hosting.
• Volume is high and predictable. Coding agents, document processing and long-running agent workflows use a lot of tokens, so cost per task matters more than cost per seat. The coming ChatGPT usage limits, which from December put ChatGPT and Codex on one shared token allowance, show how quickly heavy agent work eats into a seat-based plan.
• You need the model to stay put. Open weights mean the version you tested is the version you run, and access cannot be withdrawn by a provider mid-project.

Whichever model you pick, the architecture matters more than the headline. Keep a model-agnostic layer so you can swap between open-weight and closed models such as GPT-6.1 Terra as prices and capability change, which is the core of good ChatGPT and LLM integration. Ground answers in your own approved documents with RAG knowledge assistants, and give AI agents only the permissions they need, because a capable model can still be misled by the content it reads. If you are still comparing tools, our free choose AI tools for business course is a good place to start, and our custom AI development team can run a side-by-side test on your real tasks.

This is general information based on Mistral’s announcement, model card and licence, Hugging Face’s release notes and published media reporting as at 2 November 2026. Benchmark results are reported by the vendors. It is not legal or licensing advice; check the licence terms against your own use case before deploying.

Featured image: Photo by panumas nikhomkhai on Pexels.