Mistral's open-weight line now includes Mistral Large 3, Small 4, and the Ministral 3 series (14B/8B/3B), all Apache-2.0, alongside specialized open models like Shieldstral (safety) and Voxtral (speech). Pick a specific model and version, not "Mistral open" generally, since the lineup has grown well past the original Mistral 7B. Weights are free; you pay for the machine or a hosting provider.
Other good fits
Projects that need to inspect or modify the model itself, not just call an API.
+ Pros
Read pros
Mistral's open-weight lineup has expanded well beyond the original Mistral 7B. As of the versions checked, it includes Mistral Large 3 (v25.12) and Mistral Small 4 (v26.03), both Apache-2.0, plus the Ministral 3 series in three sizes (14B, 8B, 3B), also Apache-2.0. Specialized open models exist too: Shieldstral (a safety/moderation-focused model) and Voxtral (speech: Mini Transcribe Realtime and Small), both Apache-2.0. All of these are weights you can download, inspect and self-host, unlike Mistral's proprietary Medium and Large-tier API-only releases.
Pick it when…
You want to self-host and need real size options: Ministral 3B for the lightest hardware, up through Large 3 for the most capability.
Apache-2.0 licensing matters for your project, with unrestricted research and individual use built in.
You need a specialized open model (safety moderation via Shieldstral, speech via Voxtral) rather than a general chat model.
− Cons
Read cons
Smaller models in this line (Ministral 3B, 8B) suit scoped, well-defined tasks; test harder reasoning or open-ended tasks specifically before relying on a small model for them, rather than assuming size doesn't matter. Running Large 3 yourself needs real infrastructure investment; if you don't have that, a hosted API (Mistral's own Medium/Large tier, or another provider) may be more practical than trying to self-host the largest open option.
Pick something else when…
You want Mistral's strongest available model and don't need to self-host.Mistral Medium 3.5 is the proprietary, API-only option Mistral positions above parts of this open line for general tasks and coding.
You'd rather not manage hosting for a large model.OpenRouter can serve Mistral's open models without you running the infrastructure.
You want the simplest local runtime for whichever size you pick.Ollama is a straightforward way to run any of these locally.
Avoid if
You can't verify hardware needs for the specific size you pick; larger open models still need real GPU resources.
Context
"Mistral open" now spans a real range of sizes and specializations, not one model; a mention of "Mistral 7B" in older material refers to a since-superseded release. Always confirm the exact current model name and version (Large 3, Small 4, Ministral 3 14B/8B/3B, Shieldstral 1.0, Voxtral Mini/Small) before deciding what fits your hardware and task.
Requirements
Hardware needs scale enormously across this line: Ministral 3B is light enough for modest local hardware, while Large 3 needs serious GPU resources comparable to other large open models. This review doesn't establish exact memory or throughput figures for any specific size; test your target model on your actual hardware before committing.
Licence & cost
All weights in this line are free to download under Apache-2.0. You pay for the machine: local hardware if you self-host, or a hosting provider's compute if you'd rather not manage GPUs yourself. Apache-2.0 covers research and individual use without restriction; check Mistral's terms for any conditions on large-scale commercial deployment before assuming zero obligations at scale, since commercial terms for derivatives and production use have historically been treated separately from the open license for research use.
First task
Documentation-based suggestion — not run here.
Pick the smallest model that handles your task, starting from Ministral.
Download Ministral 3 8B (a reasonable middle size to start testing with) through a runtime like Ollama.
Write three representative prompts from your actual use case (not a generic benchmark question).
Run all three and judge whether the answers are good enough for your task.
If not, step up to Ministral 14B or Small 4 and repeat, rather than jumping straight to Large 3.
Done when you've found the smallest model in the line that handles your three representative prompts acceptably, since that's the cheapest one to run in production.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Mistral's models overview and pricing documentation, checked 23 September 2026. No model was downloaded, installed or run for this review.
Mistral AI · Sep 23, 2026 — Company overview.
Mistral models overview · Sep 23, 2026 — Full open-weight line: Large 3, Small 4, Ministral 3 (14B/8B/3B), Shieldstral 1.0, Voxtral Mini/Small, all Apache-2.0, with versions.
Mistral pricing · Sep 23, 2026 — Apache-2.0 for research/individual use vs. separate commercial license terms for derivatives and production use.
Further review notes & alternatives
Alternatives and trade-offs
Mistral Medium 3.5 wins when you want Mistral's recommended general-purpose model and don't need to self-host.
Ollama wins as the simplest local runtime for whichever size in this line fits your hardware.
OpenRouter wins when you want these models served for you rather than running your own infrastructure.
Run small open-weight models on laptops or edge hardware.
Process images and long context locally when configured accordingly.
Choose an Apache-licensed coding sibling for software work.