Large 3, Medium 3.1, and Ministral 3

The Other Pillars of the General-Purpose Lineup

Mistral Small 4 is the default recommendation, but a default isn’t a universal rule. Two situations challenge it, found at opposite ends of the spectrum: tasks requiring maximum raw power, and those that must run on-device without ever touching the network. This is precisely the space covered by the other three general-purpose models in the lineup, which we’ll cover in this order.

Mistral Large 3 — The Multimodal Frontier

Launched in December 2025, Mistral Large 3 is the most powerful model in the ecosystem. It has 675 billion parameters total, with only 41 billion activated for each token, thanks to the MoE architecture already seen with Small 4—a much more generous activation factor, which explains both the quality and the cost. It is natively multimodal: text and images enter the same request, without using a separate image analysis service. And it belongs to the Open tier, meaning its weights are published on Hugging Face.

You’ll choose it for complex multi-step reasoning, where every detail counts and an approximation early in the chain invalidates the entire conclusion. You’ll also choose it when analysis covers both text and images jointly—interpreting a chart, diagnosing an interface screenshot, translating a technical diagram into a specification. More broadly, it stands out when absolute quality takes precedence over budget, as well as in research and experimentation work where latency matters little.

The trade-off is clear: inference costs are higher and responses arrive more slowly than with Small 4. For most production applications, Small 4 remains the best compromise, and Large 3 is reserved for the fractions of the flow that truly justify it.

Mistral Medium 3.1 — The Frontier Compromise

Released in August 2025, Mistral Medium 3.1 occupies the intermediate position. It is a frontier multimodal model in the Premier tier, accessible only via the API: its weights are not published, so you cannot host it yourself.

This positioning addresses a fairly specific enterprise profile. These are teams that find Small 4 a bit short on their most complex tasks, but for whom Large 3’s latency is incompatible with interactive production use. They need reliable multimodal support in real-world conditions, and above all a model fully managed by Mistral: no GPUs to provision, no version upgrades to orchestrate. If your main constraint is delivering quickly without building an infrastructure team, Medium 3.1 is the lineup’s balance point.

The Ministral 3 Lineup — Edge and Embedded

Launched in December 2025, the Ministral 3 lineup addresses a growing need: running a performant model directly on the user’s device, without network calls.

It comes in three sizes. The Ministral 3 — 14B is the most capable in the edge range: it handles text and vision, rivals much heavier models on benchmarks, and finds its place on a local server or powerful workstation. The Ministral 3 — 8B is the balanced model, compact enough to run on a modern laptop while remaining capable for conversation, summarization, and classification. The Ministral 3 — 3B, finally, is the ultra-compact format intended for mobile devices and IoT; despite its size, it retains multimodal capabilities, text and vision.

The value of edge computing lies in a chain of consequences rather than a single argument. Data never leaves the device, which immediately settles the confidentiality question and considerably simplifies regulatory compliance. The response is instant since no network round-trip occurs, and the application continues to function offline. No dependency on an external cloud service remains, which is decisive for sovereignty concerns. The cost, finally, changes: once the model is deployed, there is no more usage-based billing. A technician working on-site without network coverage illustrates all these points at once.

All three Ministrals are available with open weights, in the Open tier. For simple local use, Ollama and llama.cpp are enough to get started in minutes; for optimized serving in production, you’ll turn to vLLM or TensorRT-LLM. Both safetensors and GGUF formats are published on Hugging Face, the latter being the one consumed by lightweight runtimes.

Summary: Which General-Purpose Model to Choose

Key Takeaways

• Mistral Large 3 (675B/41B active) is the most powerful multimodal frontier model, open source
• Mistral Medium 3.1 is a frontier compromise accessible only via API (Premier tier)
• Ministral 3 comes in three sizes (3B, 8B, 14B) for edge and embedded deployment
• The choice between these models depends on the triad quality / cost / deployment constraints
• All Open models are available on Hugging Face for self-hosting