What Do You Need to Run Mistral Locally? VRAM by Version
Mistral currently has 5 versions you can deploy locally. The smallest, Mistral 7B v0.3, needs 7GB of VRAM at 4-bit; the largest, Mistral Small 3.2 24B, needs 17GB. The cheapest device that gets one running is the Tesla V100 16GB, at about $370.
| Version | Released | Context | License | ||
|---|---|---|---|---|---|
| Ministral 3 14BDense | Dec 2, 2025 | 13.5B | 13.5B | 256K | Apache-2.0 |
| Ministral 3 8BDense | Dec 2, 2025 | 8.4B | 8.4B | 256K | Apache-2.0 |
| Mistral Small 3.2 24BDense | Jun 19, 2025 | 24B | 24B | 128K | Apache-2.0 |
| Mistral NemoDense | Jul 17, 2024 | 12.3B | 12.3B | 128K | Apache-2.0 |
| Mistral 7B v0.3Dense | May 22, 2024 | 7.3B | 7.3B | 32K | Apache-2.0 |
About Mistral
Mistral is the model family from Mistral AI, based in France. The company was founded in Paris in April 2023 by three founders who came from Google DeepMind and Meta. Its first model, Mistral 7B, released that September, already came with open weights under Apache 2.0. The lineup has since grown from small on-device models all the way to MoE flagships, alongside closed commercial models.
Mixtral 8x7B, from December 2023, is a sparse MoE model, and Hugging Face said it set a new best for open models at the time. In July 2024, Mistral AI and NVIDIA released the 12B Mistral Nemo together. The licensing went back and forth along the way: in 2024, Codestral, Mistral Large 2 and others moved to licenses that restrict commercial use, and when Mistral Small 3 came out in January 2025, the company said its general-purpose models were returning to Apache 2.0. The Mistral 3 series, released in December 2025, includes the MoE flagship Mistral Large 3, with 675 billion parameters (41 billion active per token), and three Ministral 3 models at 3B, 8B and 14B; all of them can read images, and all are Apache 2.0. Mistral Small 4, from March 2026, combined reasoning, image understanding and coding in a single 119-billion-parameter MoE.
- VendorMistral AI
- Versions5
- Parameters7.3B – 24B
- Latest versionMinistral 3 8B
- License
- Apache-2.0
- Official site
- mistral.ai
- Hugging Face org
- mistralai
How much VRAM do Mistral models need at 2-bit?
2-bit, also called Q2, covers common GGUF files such as Q2_K and UD-Q2_K_XL. They are the smallest and lose the most quality, which makes them a way to get a model running first when VRAM is tight. Each row in the table below is a Mistral version and each column is a context length; each cell is the total VRAM needed to deploy it at Q2 (weights + KV cache + 1.1GB runtime overhead).
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Ministral 3 14B13.5B | 5.5GGUF | 11.9 | 28 | 49.5 | — |
| Ministral 3 8B8.4B | 3.6GGUF | 9.2 | 22.9 | 41.2 | — |
| Mistral Small 3.2 24B24B | 9.5GGUF | 15.9 | 32 | — | — |
| Mistral Nemo12.3B | 5.4GGUF | 11.8 | 27.9 | — | — |
| Mistral 7B v0.37.3B | 2.5GGUF | 7.9 | — | — | — |
How much VRAM do Mistral models need at 4-bit?
4-bit, also called Q4, corresponds to the widely used Q4_K_M files, which balance size and quality and are the default choice for local deployment. Each row in the table below is a Mistral version and each column is a context length; each cell is the total VRAM needed to deploy it at Q4.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Ministral 3 14B13.5B | 8.2GGUF | 14.6 | 30.7 | 52.2 | — |
| Ministral 3 8B8.4B | 5.2GGUF | 10.8 | 24.5 | 42.8 | — |
| Mistral Small 3.2 24B24B | 14.3GGUF | 20.7 | 36.8 | — | — |
| Mistral Nemo12.3B | 7.5GGUF | 13.9 | 30 | — | — |
| Mistral 7B v0.37.3B | 4.4GGUF | 9.8 | — | — | — |
How much VRAM do Mistral models need at 8-bit?
8-bit, also called Q8, corresponds to Q8_0 files, which come closest to the original in quality and are also the largest. Each row in the table below is a Mistral version and each column is a context length; each cell is the total VRAM needed to deploy it at Q8.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Ministral 3 14B13.5B | 14.4GGUF | 20.8 | 36.9 | 58.4 | — |
| Ministral 3 8B8.4B | 9GGUF | 14.6 | 28.3 | 46.6 | — |
| Mistral Small 3.2 24B24B | 25.1GGUF | 31.5 | 47.6 | — | — |
| Mistral Nemo12.3B | 13GGUF | 19.4 | 35.5 | — | — |
| Mistral 7B v0.37.3B | 7.7GGUF | 13.1 | — | — | — |
FAQ
Is Mistral open source and free for commercial use?
Yes. All 5 versions listed on this site (Mistral 7B v0.3, Mistral Nemo, Ministral 3 8B, Ministral 3 14B and Mistral Small 3.2 24B) are licensed under Apache 2.0, so you can use them commercially, fine-tune them and redistribute them, as long as you include the license, keep the copyright notices and mark any files you changed. The one that's easy to mix up is the 2410 release of Ministral 8B. It uses the Mistral Research License, which allows only non-commercial uses such as personal or research work, and commercial use requires a separate license from Mistral. It's the 2512 release, Ministral 3, that is Apache 2.0, so when you download, check that the repository name says Ministral-3.
Can you run Mistral with Ollama?
Yes. All five versions have official tags: mistral:7b (which is v0.3), mistral-nemo:12b, ministral-3:8b, ministral-3:14b and mistral-small3.2:24b. Note that mistral-small:24b points to the older Mistral Small 3 (the 2501 release), not 3.2. Version 3.2 lives in the separate mistral-small3.2 library, and it can read images. Ministral 3 needs Ollama 0.13.1 or later.
