You can run Mistral locally today with four model lines. Ministral 3 (3B, 8B, 14B) and Devstral Small 2 run on a normal PC with a graphics card or on a Mac, Mistral Small 4 with 119 billion parameters needs a machine with a lot of memory such as the DGX Spark, and Mistral Medium 3.5 with 128 billion parameters belongs on a workstation or a server. Ministral, Devstral Small 2 and Small 4 come under Apache 2.0 and are free to use commercially. Medium 3.5 has a modified MIT license that excludes companies with more than 20 million US dollars in monthly revenue. Mistral launched its flagship Large 4, with about one trillion parameters, as a preview on October 6, 2026 and released the weights on October 29, 2026. It comes under the Mistral Ecosystem Open-Weight License, a modified MIT license: commercial use is allowed, and only those who offer the model to third parties as a service need a separate license from Mistral. Large 4 does not run on a single machine; the NVFP4 version alone is 610 GB.
I run my business with a team of fifteen AI employees built on Claude and test local models on a DGX Spark. I did not run a Mistral model myself for this article. I also did not measure Mistral Small 4 in my DGX test series, because it does not fit under my memory rule. What you read here comes from Mistral's model cards, the Large 4 announcements, Mistral's docs and the Ollama library, retrieved on October 26, 2026 and checked again on November 15, 2026. Everything I write about local AI is collected on my AI page.
Mistral models at a glance: what runs locally
For getting started on an ordinary machine, Ministral 3 and Devstral Small 2 are the candidates. Mistral says Ministral is built for "edge deployment", meaning small devices, while Devstral Small 2 is for coding agents that search code and edit several files. According to its card, Mistral Small 4 merges the previous lines for chat, reasoning and coding into one model and understands German alongside dozens of other languages. It activates only 6.5 billion parameters per token, but it still needs memory for all 119 billion.
The licenses: Apache 2.0 and one revenue cap
I'm a developer, not a lawyer. This isn't legal advice, just the operator's perspective: my reading of the license texts.
Apache 2.0 applies to Ministral 3, Devstral Small 2, Mistral Small 4 and Mistral Large 3. Mistral writes, for example on the Ministral 3 card: "Open-source license allowing usage and modification for both commercial and non-commercial purposes." For a small business using a model on its own network or in a product for clients, that means no revenue cap.
Modified MIT applies to Mistral Medium 3.5. The key sentence in the license reads: "You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million (or its equivalent in another currency) for the preceding month." Above that, the license says you can ask Mistral for a commercial license or use the model through Mistral's services. For most freelancers and small companies the cap does not apply.
Mistral Ecosystem Open-Weight License applies to Mistral Large 4. It is also a modified MIT license, but without a revenue cap. The key condition reads: "If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose." Under the license, Model as a Service means giving third parties access to the model through an API or hosted endpoint in a way that lets them meaningfully control inputs, parameters or training data. Internal use and products that only embed the model in specific features are exempt. If you use Large 4 inside your own company or embedded in a feature of your product, the condition does not apply by my reading.
How the Mistral licenses compare with Llama, Qwen or DeepSeek is covered in my article on open-weight licenses.
Why Mistral matters for businesses in Germany and the EU
Mistral is a European provider and opened a hub in Munich on September 28, 2026. In the announcement, Mistral writes about its open weights: "Our models run on the customer's own infrastructure, trained on their data, operated under European law, with full auditability and no data leaving the organization." That only holds if you actually run the model yourself. If you use Mistral through the API or Le Chat, it is a cloud service like any other, with a contract, data processing terms and all the questions that come with them.
For Large 4, Mistral also announces a deployment in Europe "that Mistral operates end-to-end, independently of other digital service providers and under European law". That too concerns the API, not local operation. I describe the privacy difference between your own server and the cloud in Local AI and Privacy: Cloud or Server?.
Getting Mistral running locally: Ollama, llama.cpp, vLLM
Ollama is the easiest route. Ministral 3 and Devstral Small 2 are in the library and run with one command:
ollama run ministral-3:14b
ollama run devstral-small-2
Mistral Small 4 still had no entry in the Ollama library on November 15, 2026. On the card Mistral recommends vLLM and points to GGUF files from Unsloth for llama.cpp. SGLang is listed too. The example on the card starts the model in vLLM across two GPUs with the full context of 262,144 tokens. There is also an NVFP4 version of 70.8 GB for current Nvidia hardware.
Mistral Medium 3.5 is in Ollama, at 80 GB in 4 bit. The card carries a note worth knowing: the Transformers configuration initially had a wrong entry that degraded performance on long context. According to Mistral, GGUF files generated before the fix are affected too. So use a current version.
How to set up the tools yourself, from a Mac with Ollama to a server with vLLM and SGLang, is covered in Setting Up Local AI: Ollama to SGLang. Third-party measurements of Mistral Small 4 on the DGX Spark are in Best LLMs for the DGX Spark, Measured.
Mistral Large 4: what is known and what is missing
Mistral launched Large 4 as a public preview on October 6, 2026. According to the announcement it has one trillion parameters, 52 billion of them active, handles text and images, and was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own data centers in Europe. Mistral says the training data spans more than 160 languages, including every official language of the EU. Since November 5, 2026, Large 4 has been generally available in Mistral's API, at 0.70 US dollars per million input tokens, 0.07 US dollars for cached input and 2.10 US dollars per million output tokens, according to Mistral. In the preview it was 1.36 and 4.18 US dollars.
Mistral released the weights on Hugging Face on October 29, 2026, as an improved version after three more weeks of reinforcement learning. There is a main version of 1,061 GB and an NVFP4 version of 610 GB aimed at Nvidia's Blackwell GPUs B200, B300, GB200 and GB300. Mistral recommends vLLM and also lists SGLang, Transformers and llama.cpp. The example commands on the card spread the model across 8 GPUs. In Ollama, Large 4 still exists only as a cloud model on November 15, 2026.
Even with open weights, Large 4 is not a model for your desk. The NVFP4 version needs 610 GB for the weights alone, more than the roughly 500 GB that one trillion parameters in pure 4 bit would work out to, because Mistral keeps part of the layers in FP8. Nvidia quotes models of up to 700 billion parameters for four connected DGX Spark units with 512 GB together. For comparison: for its predecessor Large 3 with 675 billion parameters, Mistral lists a whole node of B200 or H200 GPUs. For a business, "Large 4 locally" therefore means your own server in a data center or rented GPUs.
Frequently asked questions
Can I use Mistral locally for free?
Yes, Ministral 3, Devstral Small 2 and Mistral Small 4 come under Apache 2.0 and have no license fee, not even for commercial use. You only pay for hardware and power. For Medium 3.5 the revenue cap of 20 million US dollars per month applies.
Which Mistral model runs on a normal PC or Mac?
Ministral 3 in 8B or 14B, and Devstral Small 2. For Devstral Small 2 Mistral names a single RTX 4090 or a Mac with 32 GB of memory; in Ollama it is 15 GB. According to Mistral, Ministral 3 14B fits into 24 GB of GPU memory in FP8.
Is Mistral Large 4 open source?
Mistral released the weights on October 29, 2026, under the Mistral Ecosystem Open-Weight License, a modified MIT license. Mistral itself calls it an open-weight model. You may use and modify Large 4 commercially; if you offer it to third parties as a service through an API, you need a separate license from Mistral.
Does Mistral Small 4 run in Ollama?
On November 15, 2026, the official Ollama library had no entry for Mistral Small 4. Mistral recommends vLLM and points to GGUF files from Unsloth for llama.cpp.
Is Mistral more privacy-friendly than a US provider?
Run locally, no text leaves your network, no matter who made the model. Through Mistral's API it is a cloud service for which, as with any provider, you need to check the contract and data processing terms. For Large 4 Mistral announces its own deployment in Europe. This is not legal advice.
Where to go from here
To get started on your own machine, ollama run ministral-3:14b is enough; the steps are in Setting Up Local AI: Ollama to SGLang. If you are planning your own hardware for Small 4 or Medium 3.5, Your Own AI Server: Hardware and Cost and AI Server for Small Businesses will help. For an overview of all open models, see my comparison of open source AI models. Which models in Ollama run locally and which only in the cloud is covered in Ollama Models: Local or Cloud Only?.
